# Built with Opus 4.7: a Claude Code hackathon: Project Gallery - **Event:** [Built with Opus 4.7: a Claude Code hackathon](https://cerebralvalley.ai/e/built-with-4-7-hackathon) - **When:** Apr 21 at 12:00 PM – Apr 27 at 2:00 AM (EDT) - **Where:** Online - **Hosts:** [Cerebral Valley](https://cerebralvalley.ai/u/cv), [Anthropic](https://cerebralvalley.ai/u/anthropicai) - **Projects:** 288 (9 placed) - **Page:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery ## Projects ### 1. Through the flames I built it for my grandmother because she is bad with smartphones Elderly people lose billions to scams yearly, panic during emergencies, and struggle with basic apps. Bigger fonts aren't the answer. HelpGranny doesn't teach grandma to use the phone, it uses the phone for her. She just talks. The AI taps, scrolls, types, and navigates like a grandkid sitting next to her. -Scam Shield — spots sketchy calls/texts and warns her. Out loud. -Emergency — one sentence, call placed to emergency contact -WhatsApp — reads messages, sends messages, all by voice. No UI to learn. Phone just listens and acts. - **Team:** [Rafay mustafa](https://cerebralvalley.ai/u/Rafay) - **GitHub:** https://github.com/rafaym1/HelpGranny - **Demo video:** https://youtube.com/shorts/Xud7VD3zEgg?feature=share - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=21 ### 2. Maciek Regula is a self-driving NIS2 compliance audit on Claude Opus 4.7. A 9-agent pipeline (two Managed Agents + Extended Thinking) runs a plain-language interview, simulates a Red Team audit, and ships 5 remediation-ready PDFs in 15 minutes. Article 21(2) mandates the exact ten cybersecurity controls that stop real attacks — Regula identifies your gaps for ~160,000 EU companies in direct scope plus a much larger supply-chain tail. Compliance is the legal frame; cybersecurity uplift is the actual outcome. - **Team:** [Maciek Lagwa](https://cerebralvalley.ai/u/Maciek1488) - **GitHub:** https://github.com/posgame3/regula - **Demo video:** https://www.youtube.com/watch?v=TeDq3tFwJiM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=39 ### 3. Tortilya In five days, one researcher and Claude Opus 4.7 broke the most capable LLMs on the planet — twenty-four different ways — then turned every failure into a lesson, a defense, and a public certification course. This project started as adversarial research, then evolved into defensive playbooks, and finished as an enterprise learning & certification platform — all targeting one question: How can organisations stay safe in this new world full of AI agents? - **Team:** [Ilja Nevolin](https://cerebralvalley.ai/u/tortilya) - **GitHub:** https://github.com/inevolin/agentic-ai-safety-and-security-program - **Demo video:** https://www.youtube.com/watch?v=X1Zh8RTZdDY - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=80 ### 4. Rayquaza 1 in 6 adults worldwide care for an aging or chronically ill relative — managing their loved one's healthcare across a stack of paper, three portals, and four specialists who never talk to each other. Drug interactions go unflagged. Kidney trends are spotted six months late. Appeals expire silently. CareCoord is a five-agent medical document coordinator built entirely on Claude Opus 4.7's Managed Agents API. Photograph anything — pill bottles, lab printouts, insurance denial letters, appointment cards, in any language — and five specialist agents (Medication, Insurance, Medical, Appointments, Orchestrator) analyze in parallel. The orchestrator finds the cross-domain connections no single specialist sees: the kidney decline + the diuretic-NSAID pair + the denied claim still inside its appeal window are ONE event, not three. From that single analysis the engine renders five distinct documents from structured tool calls — caregiver brief (grade 6), SBAR clinical handoff, insurance appeal, attorney evidence pack, bedside note (grade 4) — re-rendered instantly when you toggle audience. Every load-bearing Opus 4.7 Managed Agents primitive ships in production: agents, environments, sessions, custom tools, memory filesystem, outcomes, xhigh thinking, permission policies, prompt caching, SSE streaming. 417 tests across 35 files, zero live API calls. Demo Mode runs in 60 seconds with $0 LLM spend. - **Team:** [Sadiq Khan](https://cerebralvalley.ai/u/sadiqkhzn) - **GitHub:** https://github.com/sadiqkhzn/CareCoord - **Demo video:** https://youtu.be/-4nUkKI2aaM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=120 ### 5. OCTAVERSE 1 in 5 businesses worldwide fail because of location. Not the idea. Not the execution. The street they picked. Locus is an AI system that investigates a location before you sign the lease. Drop a pin, describe your business idea, and 8 specialized agents built on Claude Opus 4.7 go to work using the Managed Agents API. The Cartographer pulls competitor and places data. The Observer fetches real Street View images and uses Claude Vision to count foot traffic and read storefronts. The Ethnographer searches the web for who actually lives and spends money there. The Economist pulls live weather, income, and rent data to stress-test the numbers. The Historian reads real customer reviews to find needs nobody has met. The Orchestrator connects everything and drafts a verdict. Then the Adversary attacks it, re-dispatching any of the other agents mid-investigation to dig deeper into specific gaps. The Judge delivers the final answer, every risk named. And if the location is wrong, Locus finds you a better street. Validated against 20 real-world cases with known outcomes. 17 of them were correct. - **Team:** [chedhly ghorbel](https://cerebralvalley.ai/u/OCTA) - **GitHub:** https://github.com/cheedli/OPUS-4.7-HACKATHON - **Demo video:** https://youtu.be/9VrjTZzjxvE - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=181 ### 6. Berlayar Aether is a canvas that acts: an AI-native creative system for turning one intent into multi-format, multimodal social campaigns. Creative directors and creators today jump between multiple tools, losing context at every handoff. Even after a key visual lands, they still pay the “format tax”: resizing, rewriting, recropping, localising and rebuilding the same idea across placements and languages. Aether rethinks that workflow for what’s next. Drop in references, prompts, URLs or files, and Opus 4.7 turns them into meaningful creative components: brand context, research signals, visual clusters, copy, safe zones, key visuals and variants. The canvas stays editable, with deterministic controls like alignment guides, segmentation, transparent elements, smart text placement and global/local edit propagation. Claude can plan, use tools and coordinate agents, while the creator remains in control. The goal is not “Canva with AI.” It is a responsive creative substrate for social production. - **Team:** [Enjiao Chen](https://cerebralvalley.ai/u/erniesg) - **GitHub:** https://github.com/erniesg/aether - **Demo video:** https://youtu.be/c2h9tQMahrA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=222 ### 7. Semantic GPS Semantic GPS is the governance gateway between AI agents and the business systems they were never supposed to touch unsupervised. It sits between any MCP-connected agent and any vendor MCP server as one control plane: shadow → enforce live policy mode swap, audit trail on every call with grouped trace IDs, saga rollback with explicit per-step input mapping, and a Tool Relationship (TRel) MCP extension so agents can discover safe workflow paths instead of guessing. Twelve gateway-native policies span seven governance dimensions including time gates, identity, residency, data hygiene, and idempotency. A side-by-side Playground proves the contrast between raw and governed under identical Opus 4.7 (Or bring your own LLM via APIkey) prompts. Three of the EU AI Act's hardest operational obligations: - Article 9 (risk management), - Article 12 (record-keeping) - Article 14 (human oversight) map directly onto primitives Semantic GPS tries to ship, ahead of the August 2, 2026 enforcement date. The recent Replit, Meta, and Cursor incidents all happened at the same seam: agent reaches business system unsupervised. Semantic GPS could become the missing layer. - **Team:** [Mihael Bosnjak](https://cerebralvalley.ai/u/Mboss37) - **GitHub:** https://github.com/mboss37/semantic-gps-hackathon - **Demo video:** https://www.youtube.com/watch?v=fYh2MpMw1ng&feature=youtu.be - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=223 ### 8. ORIGAMI TECH Walk the warehouse, Claude does the audit. WaybillAgent is filmed on location at a working Mydawa pharmacy godown in Nairobi. A field worker walks an aisle, points an iPhone at each shelf, and Claude Opus 4.7 reads both the bin location label and the product packaging in a single pass — including torn, faded, shadow-occluded labels that handheld Zebra/Honeywell scanners often fail to read. The system reconciles observations against Microsoft Business Central master data and returns a variance report with a per-scan citation chain. This is built for East African enterprise warehouse reality: audits that currently take two weeks, three people, and still produce paper trails with weak provenance. WaybillAgent compresses that into a 40-minute phone walk with auditor-grade evidence and explainable output. In practice, it shifts teams from specialized scanner hardware to the phone already in every supervisor’s pocket, while improving traceability and confidence for finance and operations. - **Team:** [Joseph Rwanda](https://cerebralvalley.ai/u/rmjoe99) - **GitHub:** https://github.com/rmjoe99/waybill-agent - **Demo video:** https://drive.google.com/file/d/1G7AiMQMS-0r-29bhp8x95j3jAmw0Wh6Y/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=226 ### 9. repolex.ai Project subtext is a structured knowledge graph system for AI <-> Human interaction and and multiagent AI systems. It makes your AI interactions persistent, portable, and temporal so you have a complete portable history. It makes setting up a team of agents easy and fun. - **Team:** [Rob Kunkle](https://cerebralvalley.ai/u/goodlux) - **GitHub:** https://github.com/DEMOlishous/noum3na - **Demo video:** https://demolishous.github.io/subtext-www/ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=232 ### 10. roscode studios roscode is the first conversational IDE for robotics, powered by Claude Opus 4.7. It reimagines how engineers build with ROS 2 — replacing terminals, XML configs, and hours of boilerplate with natural conversation. Built during this hackathon, roscode includes three components: (1) a Python agent for ROS that operates with a confirmation gate for safe tool execution; (2) a Tauri-based Studio with packages library, agent chat, ROS graph, and integrated terminal; and (3) an in-progress VSCodium fork and extension exploring the path to a deeper IDE integration. The core thesis: robotics software is the bottleneck of the embodied AI era. As humanoids, autonomous vehicles, and industrial robots multiply, the engineers needed to program them won't scale. roscode makes ROS 2 development accessible by letting anyone — students, hobbyists, embedded engineers — build, debug, and ship robots by talking to them. This is a first draft of how robotics development looks in a few years when Claude is even more capable. Powered by Opus 4.7. - **Team:** [Ricardo Aguirre](https://cerebralvalley.ai/u/raguirre) - **GitHub:** https://github.com/raguirref/roscode - **Demo video:** https://drive.google.com/file/d/1ovIV8lhTeusLWCXWbsMKzBR22VCA2t6q/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=236 ### 11. usectx op0 is a creative studio built on ctx, a Managed Agents harness for auditable, self-modifying agents. The user briefs the studio; the agent reaches into a Zod-typed capability registry — twelve MA tools spanning composition (compose.character, compose.score, plus two distinct video paths: compose.video for Remotion-based slideshow stitching and compose.gen_video for generative clips via Replicate Veo / Kling / LTX), workspace control (customize_studio, studio_navigate, studio_run_rig), self-modification (instructions, save_rig, propose_rig), and mail (mail_list, mail_send) — and every durable change is a typed tool call with the user's reason attached. The key primitive is the rig: a reusable workflow over the same registry the live agent uses. The rig DSL supports {{steps..blobUrl}} so a video step chains off prior character/score outputs — the built-in music-video rig runs character → score → video deterministically through Vercel WDK and replays into the same audit graph the live agent populates. Capabilities flagged usableInRig are eligible; side-effect tools (mail, navigate) stay live-only. ctx's digest agent supplies the quantified proof: a coached Managed Agent improves Discord-style digest relevance from 86.0% to 93.6% in the canonical run, including an audited retire-then-upsert belief-revision moment. op0 proves the same substrate becomes a visual, musical, inbox-capable, multi-run studio — runs stack on one canvas, errors trigger one-click recovery, and the timeline editor lets you scrub trim handles to hear the cut as you make it. Same harness, richer surface - **Team:** [Arth Tyagi](https://cerebralvalley.ai/u/arth) - **GitHub:** https://github.com/arthtyagi/ctx - **Demo video:** https://youtu.be/FwuFTUBc2-c - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=267 ### 12. Sazeleo n.a - **Team:** [sachin mittal](https://cerebralvalley.ai/u/sazeleo) - **GitHub:** https://github.com/sazeleo/gemai - **Demo video:** https://www.loom.com/share/80a0cde8450e47588f685490d19a98f4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=268 ### 13. ADLC Agentic Development Lifecycle, GitOps for AI Agents ADLC is a platform that replaces traditional CI/CD pipelines with AI agents that reason, decide, and commit. You paste a GitHub repository. Five specialized Claude-powered agents analyze your code through four intelligent quality gates, Dev, QA, UAT, and Production. Every agent decision is committed to Git with a real hash, a real author, and real reasoning. The result is a deployment decision, Approved or Blocked, with a confidence score, risk signals, and a permanent audit trail that lives in your repository forever. - **Team:** [Yash Randhe](https://cerebralvalley.ai/u/yashrandhe) - **GitHub:** https://github.com/Yashu-16/adlc-platform.git - **Demo video:** https://youtu.be/KHuXB6SRPUA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=22 ### 14. Radoslav Sandov The Garden Planner is an Android (for now, based of Research of the potential users) app for small home growers and farmers. Point your phone at a garden bed, keep a harvest log, and get science-backed rotation and watering advice — all without internet or subscriptions. Works offline, no internet needed. Built for growers in Chepinci (Sofia basin, Bulgaria) but usable anywhere. Roadmap includes - **Team:** [Radoslav Sandov](https://cerebralvalley.ai/u/SNDV) - **GitHub:** https://github.com/thecharge/garden-planner - **Demo video:** https://youtube.com/shorts/a1-qU5Ytqdk?is=5ZopkHrHUROSDdzZ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=40 ### 15. Affordance Design Studio TRACE is a living PRD for Claude Code. Every engineer who uses Claude Code writes a plan, starts building, and watches the plan go stale within the hour. The PRD becomes a tombstone. The code is the only source of truth, and the why behind every tradeoff disappears into commit messages. TRACE closes that loop. A PostToolUse hook captures every Claude Code action into a per-session log. After every turn, a Stop hook fires three Opus 4.7 sub-agents in parallel — one looks for decisions, one for assumptions validated or contradicted, one for open questions. Each returns a confidence-scored proposal. Confidence ≥ 0.5 lands in a review queue. The user accepts or skips in a minimal TUI, and accepted proposals write themselves into the right section of the PRD. This repo's TRACE.md was written by TRACE itself — commit 797daaf is the moment TRACE first used TRACE to document itself. - **Team:** [Shandar Junaid](https://cerebralvalley.ai/u/Shandar) - **GitHub:** https://github.com/shandar/trace - **Demo video:** https://youtu.be/7vDUqqD3xWo - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=41 ### 16. Holo The Rapper YUME is a social simulation where Claude Managed Agents live as residents of a city. Each resident has its own independent agent session and persistent memory. No instructions are given to the agents. They receive only the state of the world — buildings, nearby people, their own health, finances, stress — and reason autonomously about what to do. This architecture makes the simulation genuinely realistic. Under the hood, YUME implements per-resident asymmetric relationship models, an independent evaluation step that scores each interaction, seed-based reproducible probabilistic events, and parallel synchronization orchestration across all agents — a structure that cannot be achieved with a single prompt. Change the city layout, swap in different residents, or alter the scenario, and an entirely different society emerges. Urban planning, policy testing, social experiments — whatever the question, YUME answers it by simulating the whole city. - **Team:** [Holo The Rapper](https://cerebralvalley.ai/u/holotherapper) - **GitHub:** https://github.com/holotherapper/yume - **Demo video:** https://youtu.be/WrKzfGpugiM?si=4_RMF7d9Un5NuqRj - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=81 ### 17. Vedant Srivastava Poneglyph is institutional memory for development consulting projects. Six specialized agents on Claude Opus 4.7 read scattered field evidence — scanned Hindi forms, meeting transcripts, CSV exports — catch silent commitment drift across stakeholder meetings, and draft pre-meeting briefings with every claim independently audited. Three Opus 4.7 capabilities exercised at production quality. Scout extracts evidence from handwritten Hindi attendance forms with pixel-coordinate bounding boxes — no OCR pipeline, no fine-tuning. Archivist uses on-demand tool reads to navigate project memory the way a human consultant uses a binder, detecting silent walk-backs the project team did not acknowledge. Auditor takes the Drafter's claims and re-reads source images independently with vision, refusing to confirm anything it cannot ground — same model, two different roles, with vision, refusing to lie. 12/12 evidence extractions, 3/3 drift detection across variance runs, 89% aggregate eval pass rate across 28 test cases. Seven documented failure modes in the repo. $200B+ flows annually from World Bank, GIZ, and UN agencies through this kind of consulting work. Built in five days. Live demo: https://poneglyph-chi.vercel.app - **Team:** [Vedant Srivastava](https://cerebralvalley.ai/u/Vedantzz) - **GitHub:** https://github.com/vedntzz/Poneglyph - **Demo video:** https://www.youtube.com/watch?v=hPqvIciYbDk - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=84 ### 18. mo_creates LoomMCP — The context compiler for AI coding agents. Every time Claude Code opens your codebase, it reads thousands of lines it doesn't need. A typical session burns 140,000+ tokens just on code reading - most of it irrelevant context. That's money wasted, latency added, and quality degraded. LoomMCP fixes this at the infrastructure layer. It's a single MCP server that indexes your codebase once using tree-sitter AST parsing, then lets Claude retrieve only the exact code it needs -- function signatures, specific implementations, references, dependencies, with byte-level precision. - **Team:** [Muhammed Nehan](https://cerebralvalley.ai/u/muhnehh) - **GitHub:** https://github.com/muhnehh/loom-mcp - **Demo video:** https://www.youtube.com/watch?v=z1gBFO4jdbI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=121 ### 19. Heimdall Hearthwatch is an AI companion for families looking after an aging parent at home. A single camera watches over Mom or Dad; computer vision tracks daily activities; Claude reasons about what's happening and only calls the family when something actually matters in the language a daughter or son would actually use rather than bombarding their phones with event notifications. Most elder-care monitoring tools detect events and beep. Hearthwatch understands situations and talks. When a fall happens, it doesn't fire a generic alarm — it looks at the saved keyframe via Claude Vision, confirms what it sees, and composes a warm alert: "Dad may have just fallen in the living room near the coffee table — call him now, and if he doesn't answer, head over right away." That specificity comes from Claude actually reading the frame. The system runs two layers in parallel. Local detectors (MediaPipe pose + YOLOv8) handle every frame on-device fall detection (including a close-to-camera vanish path for when someone falls below the field of view), pacing, inactivity, repetitive behavior, and multi-person subject tracking that follows the right person even when others enter the frame. Claude runs on a tiered cadence: Vision on keyframes to confirm and describe, Opus to reason over the rolling event window and assess risk, Haiku to compose alerts. Every answer in the chat tab is grounded in a real SQLite query — no hallucination. It ships on a Raspberry Pi 5 with a Pi Camera Module 3, costs under $150 in hardware, and runs at roughly $0.23 per monitoring session in API credits. The goal was to build the thing a real family would actually put on their kitchen counter. - **Team:** [Krishnaprasad Sreekumar Nair](https://cerebralvalley.ai/u/serious-engineer) - **GitHub:** https://github.com/serious-engineer/hearthwatch - **Demo video:** https://youtu.be/PSFa0BT7l-s - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=184 ### 20. Team Cumberland Reading time keeps ending the same way. We close the book early because we can't stand reading it. Our kids notice. We notice. And the books are everywhere — every shelf, every order from every store, sent home in every backpack. A long tail of writing that connects no one to anything. The world is dying for stories that move people, and most of what's getting made is filler. Two parents asked the obvious question: could Claude help us write the book we'd actually want to read to our own kids? Two extremes already exist and both produce a lot of mediocre books. AI writing the whole thing — slop. A human writer alone with a blank page — sometimes brilliant, often forgettable, frequently never finished. Same root cause: neither extracts the actual writer's actual voice and meaning. AI can make words, but not meaning. Meaning lives in the author. The Factory's job is to draw it out. The Book Factory is an AI-native publishing platform for authors, designed around the author rather than around the AI. The author is interviewed by the system on world, characters, voice, plot, and scene; their reference shelf and existing prose seed the style; every interview output becomes a typed contract that downstream agents must satisfy. We're starting with illustrated chapter books because that's what our first series needs. The pipeline produces any chapter book in any fiction genre. What you're seeing in this submission is the proof of concept. Stacia is an educator and dedicated reader. Dan builds AI-native production systems. We have three kids. We are the hardest possible users. We refuse to ship a book to our own kids that we wouldn't want to read. Our prototype effort is three books on Amazon by June 2026, written by us using only the Factory we're building. Eating our own cooking, with our daughter as the first-gate test reader. What we found this week: the book we're writing is better than what either of us would have produced alone. We probably never would have finished it. And the story itself is better — Claude drove the story-craft research that connected dots we wouldn't have seen. Once the proof is in, the Factory becomes a platform other authors can acces— authors with a story they can't quite finish, or who are tired of producing words that don't have meaning underneath them. The bet is bigger than three books. AI-assisted writing done right is how we raise the quality and the quantity of human-authored stories that actually move people in a world dying for them. AI can flatten meaning, and AI can amplify it. We're building the path that amplifies. Brings us closer in. And helps us make things that matter. PS - The soul of this project is a 9 year old half girl nicknamed Piña, who just moved to a new country and new school and is trying to find her way. She's thoughtful and funny. Read the first chapter or two. You won't be disappointed: https://drive.google.com/file/d/1561LhpRiRQUA9Rn6doyV1CyYvm3u4i06/view - **Team:** [Dan Cumberland](https://cerebralvalley.ai/u/dancumberland) - **GitHub:** https://github.com/dancumberland/Book_Project - **Demo video:** https://youtu.be/v_k0A9SnMZU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=224 ### 21. Just Emon Deployed on: https://web-three-kappa-0ay1jivbpz.vercel.app/ Primum is a clinical autopsy system for conversational AI failures in mental-health contexts. Tens of thousands of conversations between vulnerable users and language models are happening right now, and there is no audit pipeline for any of them. Primum is that pipeline. Drop a zip of conversations (Claude, ChatGPT, JSONL, CSV, Markdown — eight formats supported). Apollo, our forensic auditor, runs a four-wave dependency-aware analysis (frame, linguistic vitals undertones, psych profile, failure timeline, verdict), locating the precise turn where the bot failed clinically, citing the instrument that names the failure (DSM-5-TR, C-SSRS, LIWC-22, MITI 4.2.1, APA AI Health Advisory 2025). The case then routes to a peer bench of five Opus 4.7 reviewer agents — Marcus (alliance), Anika (DSM), Sam (crisis), Rina (methodology), Director Elena (veto power). Severity-4 cases skip the junior and go straight to the director. Every reviewer's reasoning is on the record. Approved cases export as a printable BMJ-format case report and a contrastive DPO/SFT training pair. The corpus that Primum produces is citable. - **Team:** [Emon Sarker](https://cerebralvalley.ai/u/emondsarker) - **GitHub:** https://github.com/emondsarker/apollo-ai-psych-warden - **Demo video:** https://youtu.be/YwykVIxEIA0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=227 ### 22. José Ángel Heras APPRA — A Curriculum Assistant for Spanish Vocational Training APPRA helps the three audiences who actually need Spain's Vocational Training (FP) curriculum — teachers, students, and families — work with it. Built end-to-end with Claude Opus 4.7. The problem. Spain's curriculum is published as dense legal PDFs in the official gazette (BOE). Each module is defined by a list of Learning Outcomes (RAs) and Evaluation Criteria (CEs). Teachers rephrase them by hand; families have no way in; students rarely meet the criteria they will be assessed against in actionable language. What APPRA does. 1. Extracts the curriculum from BOE PDFs with Claude Opus 4.7 (output_config.format: json_schema + document caching). Every RA and CE is parsed into queryable JSON — auditable line by line against the source. No paraphrasing. 2. Renders an interactive UI to browse modules, track per-criterion progress, and publish an official course state to GitHub via a Netlify Function. 3. Multi-role chatbot also on Claude Opus 4.7, streaming with prompt caching. Same model, three tones: Teacher (didactic programming, rubrics), Student (plain explanations), Family (jargon-free). Why Opus 4.7. The same model plays two roles: it prepares the canonical data and then consumes it to help end users. Adaptive thinking, structured outputs, and prompt caching over the module context make it practical and affordable. Shipped. Three real Informática modules — 20 Learning Outcomes, 157 Evaluation Criteria — extracted literally from BOE Real Decretos 686/2010 and 1629/2009. - **Team:** [Jose Angel Heras](https://cerebralvalley.ai/u/Joanh) - **GitHub:** https://github.com/joanh/APPRA - **Demo video:** https://www.flexclip.com/es/share/15764586S3YlFD7NnJvjagzCZnn37m98ICe6UOIy.html - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=23 ### 23. Inkspell by Aira Every week, content creators film, speak, and teach — but Google has no idea they exist. Their ideas stay locked inside TikTok, LinkedIn and Instagram, never indexed by search. Inkspell turns any YouTube video into a 1,000–1,200 word SEO-optimised blog post that sounds like the creator wrote it. Not AI slop — your actual voice. Built around Aira Al-Q Sinclair, an AI writing persona trained on how the creator thinks and structures ideas, Inkspell lets users describe their voice through tone chips, sentence style, loves/avoids, and a writing sample. Claude Opus 4.7 then generates a blog post with YAML frontmatter, proper H2 structure, and a concrete CTA — ready to publish. The transcript is fetched directly from YouTube. The blog post streams back in real time. Enter your API key, paste a YouTube URL, cast your spell. - **Team:** [Wan Wei Soh](https://cerebralvalley.ai/u/katamichiww) - **GitHub:** https://github.com/katamichiww/yt-to-blog-web - **Demo video:** https://youtu.be/w8xaqJuVPgg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=42 ### 24. Hathor Hathor is an autonomous clinical reasoning agent for cross-border vaccination reconciliation. A child arrives in Egypt with a Nigerian (or Sudanese, Syrian, Yemeni…) vaccination card; Hathor reads the card, resolves trade-name combinations to canonical antigens, checks every dose against the destination schedule, and emits a visit-by-visit catch-up plan — all with the agent deciding its own tool path, no hardcoded pipeline. Built on the Claude Agent SDK with eight in-process MCP tools and two safety loops: a per-field vision gate (anything < 0.85 confidence routes to clinician HITL review) and a per-recommendation rules-engine gate (a deterministic Python WHO-DAK validator vetoes any clinical claim before it reaches the UI or FHIR bundle). The clinician retains final authority and every override is logged to a FHIR Provenance resource. No open-source tool exists for autonomous cross-schedule reconciliation today; the closest prior work (AI-VaxGuide, arXiv 2507.03493) is single-country Q&A. - **Team:** [Ahmed Zayed](https://cerebralvalley.ai/u/DrZayed) - **GitHub:** https://github.com/DrAhmed7887/hathor - **Demo video:** https://youtu.be/WPEnTkUN-fI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=85 ### 25. HashtagWorld SAFER is an open-source agent control plane that turns any AI agent running on LangChain, Google ADK, AWS Strands, raw Anthropic, raw OpenAI, OpenAI Agents SDK, AWS Bedrock, or CrewAI into one observable, governable, audit-ready system. The pain it solves: every team picks a different agent framework, ships an agent, and inherits three independent gaps. There's no shared view of what the agent did (each SDK ships its own callbacks and trace IDs), no shared way to enforce policy (every team rebuilds PII redaction and tool whitelists in code), and no shared report that survives an audit. SAFER is the missing layer. A two-line adapter normalises any framework into a 10-hook lifecycle contract that streams into a self-hosted FastAPI + SQLite WAL backend. Every event flows through three engines: a deterministic Gateway that enforces PII regexes and natural-language-compiled policies before tool calls, a Multi-Persona Judge (six personas, dynamically routed, single Opus 4.7 call per hot hook with > 80% prompt-cache hit rate), and a per-step Haiku 4.5 scorer that gates expensive Opus calls. Onboarding: an Inspector built as a Claude Managed Agent statically reviews the agent's source the moment it registers running grep + bash + AST scans inside a cloud sandbox, with a shared `safer-inspector-knowledge` memory store that gets smarter every scan. Pre-deploy: a Red-Team Squad (Strategist → Attacker → Analyst, also via Managed Agents) probes the live agent against OWASP LLM Top 10. Post-run: a deterministic Python aggregator + Sonnet 4.6 Quality + Thought-Chain Reconstructor produces a Session Report and a regulator-ready Compliance Pack (GDPR / SOC 2 / OWASP LLM) in PDF, HTML, or JSON. Apache 2.0, runs entirely in your VPC, never phones home. Ships in one `docker compose up`, or as `pip install safer-sdk` + `pip install safer-backend` (PyPI v0.1.1). 401 tests passing across the SDK + backend. Hi team, A quick note I'd like to flag with my submission for full transparency. I built this project together with a teammate on GitHub; both of our commits and pushes are visible on the repository. When I went to submit the form, I learned for the first time that team members must also be among the 500 selected participants. I wasn't aware of this rule until that moment, otherwise we would have handled the registration accordingly. Sharing this proactively, and happy to follow whatever process you advise. Thanks, Emine Hanedar - **Team:** [Emine Hanedar](https://cerebralvalley.ai/u/hashtagemy) - **GitHub:** https://github.com/hashtagemy/safer - **Demo video:** https://youtu.be/kySyDsgrco4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=122 ### 26. Akash Stephen Medical students memorise twenty thousand facts. They sit through two preclinical years of flashcards, mnemonics, and question banks, and they believe — every educator believes — that this is how medicine is learned. Then they reach the bedside. A 67-year-old man collapses while climbing stairs. Not at rest. Only on exertion. The student knows the words aortic stenosis, syncope, exercise intolerance. They have seen the flashcard. And they freeze, because the patient does not look like the flashcard, and the brain they were trained to use has no machinery for what to do next. The crisis in medical education is not a knowledge shortage. It is a mental-model deficiency. We have spent a century teaching students to retrieve facts. We have never seriously tried to teach them to generate understanding from mechanism. Every existing tool — Anki, UWorld, Amboss, Osmosis, Pathoma, Sketchy — is a retrieval tool. None of them teach the student to derive. Cortex is the first tool that does. What Cortex Is Cortex is a mechanistic medical learning system. It teaches medicine the way physics is taught: not by memorising Newton's laws, but by throwing balls, watching them arc, and discovering that the formula is just compression of what you already feel. The current demo is a live cardiology tutor. The student is given a parameter-bound heart — an anatomical SVG whose chambers, valves, and great vessels are wired to six physiological knobs (heart rate, contractility, preload, afterload, systemic vascular resistance, valve compliance). When a knob moves, the visual heart responds in the same frame. When the student scribbles calcium onto the aortic leaflets, the valve area shrinks live from 3.0 cm² to 0.7 cm², the leaflets stiffen, the left ventricle strains, the lungs cloud with coral haze. Then the student picks any concept they want to learn — severe aortic stenosis, cardiogenic shock, septic shock, hypertensive emergency, the vasodilator trap, or anything else they type in — and Claude Opus 4.7 takes the wheel. It dials the sliders with smooth tweens, narrates from first principles, asks the student to predict the next move, and grades the prediction against the deterministic physics behind the visual. The student does not watch a lecture. They turn knobs on a heart, then ask Cortex to teach them whatever they do not yet understand. Each interaction with the platform creates a cognitive profile of the user's understanding of the human body and the subsequent sessions are dynamically made to teach the specific errors that the user is making. The same architecture — a parameter-bound visual world, an agentic Opus teacher with typed tools into that world, and a clinical-grounding critic refusing claims the physics does not justify — scales to teach all of medicine on a first-principles basis, backed by the way the human brain actually forms durable knowledge: through prediction, perturbation, error, and repair, not through retrieval of rehearsed facts. - **Team:** [Akash Stephen](https://cerebralvalley.ai/u/akashstephen) - **GitHub:** https://github.com/akashstephen/Cortex_Demo - **Demo video:** https://www.youtube.com/watch?v=sjnVm8lzsyk - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=123 ### 27. Operant Team Operant AI is an autonomous operations platform that turns simple instructions into fully functioning systems. Instead of manually building workflows or stitching together tools, users describe what they want to achieve, and Operant AI designs, executes, and manages the entire process. It connects to tools like Notion, Slack, and email services, transforming high-level intent into structured workflows that run in real time. What sets Operant AI apart is its self-improving execution loop. The system does more than just running tasks it monitors outcomes, detects failures, and automatically adjusts workflows to improve performance over time. This creates a shift from static automation to living systems that adapt as conditions change. Whether it’s onboarding users, managing leads, or coordinating internal operations, Operant AI acts as a continuous execution layer, reducing manual effort and enabling teams to focus on higher-level decisions rather than repetitive processes. - **Team:** [Olabowale Babalola](https://cerebralvalley.ai/u/Esimuda) - **GitHub:** https://github.com/Esimuda/OperantAI - **Demo video:** https://drive.google.com/drive/folders/1bUWG8Eki6MS8lZ-D3VHVH0oDtbKB2b6N - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=138 ### 28. AI Kitchen planner Demo: https://kitchen-planner-aiysxix6sq-ew.a.run.app/ Drag-and-drop planners accept coordinates. AI Kitchen Planner accepts goals. "Wheelchair-accessible galley with induction hob and a double-bowl sink" → a buildable 3D kitchen, modeled from real IKEA components at millimetre precision. How it works Opus 4.7 is the CAD operator, running inside an Anthropic Managed Agent. Authored skills teach it each IKEA product family on demand. The agent writes kitchen.json to its sandbox; we pull it via the Files API as a structured-output channel. A custom notify_model_updated tool drives live SSE updates, so the 3D view re-renders mid-conversation as Opus reasons through the layout. Cabinets and worktops are procedural; appliances are real GLB meshes. Sessions are resumable from the URL hash. - **Team:** [Pavel Kral](https://cerebralvalley.ai/u/pavelkral) - **GitHub:** https://github.com/pavelkraleu/anthropic-opus-hackathon - **Demo video:** https://youtu.be/Sys9pnU9Q54 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=185 ### 29. The Panel Panel is an agentic workspace for data-science experiments. Instead of one AI writing cells, four specialized Claude agents — an Implementer, Interpreter, Tagger, and Archivist — run your experiment together as a sequential jury. The Implementer plans and writes code, a Jupyter kernel executes it, the Interpreter reviews the output and flags methodological risks (leakage, collinearity, temporal splits), the Tagger classifies each step against an 8-tag vocabulary, and the Archivist decides what's worth committing to a persistent knowledge base. Every action is appended to a structured deliberation.jsonl event log. The web UI streams it live; finishing produces a shareable URL where anyone can open the experiment three weeks later and understand not just what happened, but why — the hypotheses tested, alternatives rejected, and pitfalls caught. Copilot writes cells; W&B tracks metrics; Panel captures the reasoning around them. - **Team:** [Arpan Nookala](https://cerebralvalley.ai/u/arpannookala) - **GitHub:** https://github.com/arpan1221/Panel - **Demo video:** https://loom.com/share/folder/55780b0c8e7c43eebc894c4485743a0e - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=225 ### 30. NexusLegis I'm a Peruvian cybercrime victim who has spent four years navigating a justice system with a 1.1% cybercrime conviction rate. Courts worldwide are overwhelmed — prosecutors delay, evidence goes unexamined, crimes expire under statutes of limitations. Victims and attorneys increasingly turn to AI for help drafting legal filings, and the AI hallucinates jurisprudence. My complaint citing fabricated rulings was rejected. Over 100 attorneys globally have been sanctioned for the same mistake. Legal research infrastructure is now a billion-dollar market — but none of the platforms serving Latin America's 130,000+ Peruvian attorneys tell you which citations are real and which the AI fabricated. LexGraph ingests rulings from Peru's Constitutional Court, extracts entities and relations using Claude Haiku 4.5, and answers legal research questions in Spanish or English using Claude Opus 4.7. When Opus cites a ruling the retrieval layer didn't return, the UI flags it as unverified instead of hiding it. LexGraph tells you which citations it can verify and which it cannot. Built in 7 days by a non-engineer using Claude Code. The pipeline is jurisdiction-agnostic — Peru is where I needed it first. - **Team:** [Marco](https://cerebralvalley.ai/u/marcoantonyo) - **GitHub:** https://github.com/mbarrueto/lexgraph - **Demo video:** https://drive.google.com/file/d/12npbxrPU5ssL-jSgI8O_GUBAvgk8gotO/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=239 ### 31. Lacuna Lacuna is an AI-for-science system built to say no. Most discovery tools are optimized to generate more hypotheses. Lacuna focuses on the harder scientific step: rejecting weak ones before they become convincing stories. Built with Claude Code, Opus 4.7, and Claude Managed Agents, Lacuna proposes candidate biological laws, then sends them through a fixed five-test Python gate. The AI cannot change the rules after seeing the result. Failed candidates are saved as useful scientific evidence, not treated as wasted work. On a public kidney cancer dataset with 505 patients, Lacuna rejected 194 of 203 candidate laws. One simple survivor, TOP2A − EPAS1, matched a known kidney-cancer growth program. That was the point: this is not presented as a new biological discovery, but as a methodology proof that the system can rediscover known truth under strict rules. That survivor also passed an independent survival test in a kidney cancer clinical-trial dataset. A later three-gene extension proposed by the system failed the same gate. The broader goal is a repeatable scientific discipline: propose, test, reject, remember, and only then interpret. - **Team:** [JangKeun Kim](https://cerebralvalley.ai/u/jangkeunkim) - **GitHub:** https://github.com/jang1563/lacuna-falsification - **Demo video:** https://youtu.be/eB-gREA4zGI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=240 ### 32. Yuki OpenBench OS is a protocol runtime for research and lab workflows. Labs already have approved documents like SOPs, SDS sheets, equipment manuals, and local policies, but execution at the bench is still mostly manual: operators read dense PDFs, remember safety requirements, track timing, capture deviations, and write handovers after the fact. OpenBench OS turns those documents into guided, executable workflows. It uses Opus to help compile source documents into a structured protocol graph with ordered steps, PPE, hazards, controls, timers, visual checkpoints, stop conditions, and source citations. A reviewer checks and publishes the protocol as an immutable version before it can be used. During a live run, the Bench Runtime Client guides the operator one step at a time, supports notes and timers, checks required confirmations, verifies visible conditions from bench photos, blocks unsafe or incomplete steps, logs deviations, and generates a handover report from the actual event history. The goal is to make lab procedures clearer, safer, more traceable, and easier to hand over — without replacing existing lab documents or systems. - **Team:** [Pham Ba Anh](https://cerebralvalley.ai/u/snowy_x) - **GitHub:** https://github.com/yuki-20/OS-Bench - **Demo video:** https://drive.google.com/file/d/1d289jZpylpZqkM0wNGCn3S6AuHlG8hMI/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=249 ### 33. Strata Mundo Strata Mundo — your math mastery voyage Most math tools tell you a percentage. Strata Mundo tells you how a learner reasons. A learner builds fractions by dragging brass pieces onto a target — every drag, removal, and reset is recorded as process telemetry. Claude Opus 4.7 reads the full trajectory and produces a categorical mastery map per CCSS standard: mastered, building the skill, misconception detected (with the specific named wrong mental model and traceable evidence), or not yet probed. Never percentages. A Plan Architect Managed Agent then synthesizes a tailored plan: 2–3 activities per gap, drawn from a curated multimodal library, sequenced concrete → representational → abstract. The library grows from contributions. Anyone — guides, parents, teachers, learners themselves — can submit a new activity. Every submission is reviewed by Opus 4.7 against six published pedagogical criteria, then by a human. Criteria are public on the methodology page. Built solo in 4 days for the Cerebral Valley Built with Opus 4.7 hackathon. Live at stratamundo.com. Everything user-facing — home, methodology, contribute, search, voyage, report — uses Opus 4.7 as the substantive reasoning engine. Categorical, never numeric. Misconception-named, never vague. Community-grown, AI-vetted, human-approved. - **Team:** [Barbara Jauregui Wurst](https://cerebralvalley.ai/u/Dadababa) - **GitHub:** https://github.com/barbarajauregui-dadababa/stratamundo - **Demo video:** https://youtu.be/EVB_bCUYySg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=250 ### 34. eBBD Paper Trail is a Claude-powered agent and web dashboard that investigates why an ML paper's published result fails to reproduce from its public repo. It identifies the root cause, applies the minimal fix, re-runs the evaluations, and opens a real GitHub PR with an evidence-backed scientific dossier. It targets the boring, expensive layer of the reproducibility crisis: data leakage, train/test contamination, label-aware preprocessing, and metric-implementation drift; the failures that quietly inflate headline numbers and rarely get caught in review. The product runs in two modes. Deep Investigation is the autonomous flow: paste a paper URL and a repo URL then watch the agent generate ranked hypotheses, run discriminating checks via tool use, converge on a verdict with cited evidence, patch the code, and post a before/after metric delta. Quick Check is a chat-style sidebar where a researcher asks targeted, bounded questions ("is this imputation fit on train only?") and gets a verdict (confirmed OR refuted OR unclear) with at least one code citation in under 30 seconds. Together they reframe the agent from "autonomous scientist" to "verification intern"; one a researcher can actually trust to do the unglamorous audit work. - **Team:** [Enam Biswas](https://cerebralvalley.ai/u/ebis) - **GitHub:** https://github.com/e-biswas/paper-trail - **Demo video:** https://youtu.be/j_11iV3FKdw - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=24 ### 35. Kyōjo (共助) Kyōjo (共助 — "mutual aid") is a real-time humanitarian coordination platform for Japan's disaster response ecosystem. When a major earthquake or typhoon strikes, the gap between what people need and what reaches them is not a technology problem — it is a coordination problem. Emergency operators juggle disconnected systems. Foreign residents fall off the edge of Japanese-only alerts. Volunteer capacity sits idle while open needs go unfilled. Kyōjo solves this with three interlocking capabilities: 1. Interactive Japan Map — A Canvas 2D Mercator map centred on Japan shows aid-network flows between 8 cities across 8 regions. When a disaster scenario is selected (Nankai Trough, 2024 Noto, Typhoon, 2011 Tōhoku, or custom What-if), arc paths react in real time: gold flows break into dim disrupted lines and coloured emergency routes appear at the epicentre. 2. Claude Opus 4.7 Orchestrator — Dispatches 10 specialised subagents in 3 dependency-ordered waves to produce a complete, citation-verified operations packet in under 15 minutes: seismic situation report, damage assessment, real-time shelter discovery via Google Places, population vulnerability overlay, resource matching, and multilingual public alerts (JA + EN + dynamic third language via TTS). 3. Living Knowledge Base — 22-document preparedness corpus (Cabinet Office whitepapers, JMA guides, prefectural hazard plans) with BM25 retrieval, an AI chat interface, and enforced source citations — available in 14 languages. Built for NGOs, volunteer centres, and government responders. Fully accessible (WCAG 2.2 AA, JIS X 8341-3, reduced motion, audio alerts). Zero invented facts: the Verifier agent blocks packet release if any claim cannot be traced to a source. - **Team:** [pascal burume](https://cerebralvalley.ai/u/pascal_b) - **GitHub:** https://github.com/PascalBurume/kyojo - **Demo video:** https://www.youtube.com/watch?v=E8-a9TNSjlY - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=43 ### 36. Pablo Araujo Of about 500 developers approved for the Cerebral Valley "Built with Opus 4.7" hackathon, only 2 are Brazilian. The same ratio shows up across fellowships, grants, and accelerators. Brazil has over a million developers. The talent isn't missing. It's unmapped. Radar is a career plan, not a job board. Three Claude Opus 4.7 agents working together. Anamnesis reads you, your GitHub, your CV, your trajectory, and writes a profile honest enough to use in an application. Scout is one shared agent that runs on its own every week, crawls 1,240+ curated sources (FAPESP, Emergent Ventures, Chevening, MEXT, YC, Fundação Estudar, GSoC, MATS), and adds new ones it discovers along the way. Strategist matches the two: it ranks the opportunities for you specifically, writes a why-you paragraph that cites things you actually did, and lays out a 90-day plan around the cards it picked. I opened the waitlist this week. Fifty Brazilian developers signed up in five days. I sat down with five of them. An undergrad with published research and mentors who couldn't name a single fellowship he was a strong fit for. A senior mid-career engineer looking for what's next. A junior dev who got her first job through a referral and had never heard any of this existed. Different stages, same gap: not "I don't know what exists", but "I don't know what I'd be a fit for". Brazil doesn't have a brain drain. It has an information drain. Radar finds the opportunity, and finds the person. Live: radar.pabloaa.com (invite-only beta during judging) · Showcase, no signup: radar.pabloaa.com/showcase · Source: github.com/pablo-aa/radar (AGPL-3.0). - **Team:** [Pablo Araujo](https://cerebralvalley.ai/u/pablo_aa) - **GitHub:** https://github.com/pablo-aa/radar - **Demo video:** https://youtu.be/ueLPzXevysQ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=86 ### 37. Juan Manuel Fraga The health system today gets paid when you get sick. Hospitals, visits, and pharmacies all run on the engine of illness. Nobody is paid to keep you well. I built Health Companion because, sitting at fifty, watching my parents age, I realized the product I needed didn't exist. It is a wellness companion whose only job is to keep people well, in the language they actually speak. Powered by Opus 4.7 — extended thinking exposed as an opt-in clinical audit layer any user can open, multimodal reading of lab reports, prescription labels, and smartwatch screens with no OCR in between, and a living state document that compounds with every conversation. It never diagnoses, never prescribes, always refers. A practicing physician — me — authors and reviews every clinical string. Where care is scarce, the value rises: someone without a family doctor now has a digital one. Especially in places like México. Most health products live around the visit. This one lives between visits — in the years when nothing is happening yet, the years that decide everything. Built in five nights with Opus 4.7 and a coordinated team of Claude Code subagents. Open source, Apache 2.0. Demo: https://health-companion-five.vercel.app/?demo=1 - **Team:** [Juan Manuel Fraga Sastrias](https://cerebralvalley.ai/u/jmfraga) - **GitHub:** https://github.com/jmfraga/health-companion - **Demo video:** https://youtu.be/-q0DTQhQW4g - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=124 ### 38. Zappy Chatalog is an AI-powered WhatsApp order management system built for Ghanaian small businesses. In Ghana and across West Africa, WhatsApp is the primary commerce channel; customers already message businesses to place orders. Chatalog formalises this: Customers order in natural language via WhatsApp, and Claude runs an agentic tool-calling loop that understands the request, checks the menu, creates the order, and replies naturally, no app download, no account needed. Store owners get a real-time admin dashboard with a Kanban board (new → confirmed → preparing → in transit → delivered), drag-and-drop status updates that instantly notify customers on WhatsApp, a natural language command bar for updating menu items, and automated payment nudges for unpaid orders. The core is an agentic loop: Claude receives a WhatsApp message, calls tools to read/write the PostgreSQL database, and iterates until it produces a final reply. Tools follow a strict read (get_*, list_*) / write (create_*, update_*, cancel_*) convention. Built with FastAPI + asyncpg on the backend and Next.js 16 + Tailwind v4 on the dashboard. - **Team:** [Emmanuel Yeboah](https://cerebralvalley.ai/u/noelzappy) - **GitHub:** https://github.com/noelzappy/chatalog - **Demo video:** https://youtu.be/jm9CYxypzC8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=135 ### 39. M-2.5 Saath is a voice-first AI medication companion for elderly Indian families, built entirely inside the Telegram chat they already use. A 68-year-old patient simply sends a photo of her doctor’s prescription, and Saath reads handwritten prescriptions in Hindi, English, or Tamil. If any field is unclear, it asks for clarification instead of guessing. Once the medication list is updated, a Claude Managed Agent runs autonomously using four custom tools: lookup_drug_info, check_pair, check_contraindication, and report_findings. It systematically checks every drug combination and every medical condition for clinically significant interactions. If no issues are found, both the patient and her daughter, often living 800 km away, receive one simple message: “No issues were found in your medicines.” If a risk is detected, such as Warfarin and Aspirin together, both phones receive a severity-graded clinical warning. The same Telegram chat also handles voice notes in Hindi, scheduled reminders with a personalized voice, smartwatch critical-alert forwarding, lab PDF parsing into plain Hindi, weekly doctor and family PDF reports, and an emergency-services regex bypass. No app installation. Just messages. - **Team:** [Kavin Thakur](https://cerebralvalley.ai/u/auraCodes), [Naman Goyal](https://cerebralvalley.ai/u/NamanGoyal) - **GitHub:** https://github.com/auraCodesKM/sath_claude - **Demo video:** https://drive.google.com/drive/folders/1lmRiyzTEk3X12Au4pBaHdIXVuErIvL1j?usp=drive_link - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=137 ### 40. Evolyth Evolyth is an autonomous AI research loop that improves machine-learning models through repeated mutation, execution, evaluation, and review. A Claude-powered coding agent proposes safe architecture changes, the system runs training jobs locally or on Google Cloud Run GPUs, then a reviewer agent analyzes the result and queues the next promising experiments. The project includes a live dashboard showing evolution progress, leaderboard, queue state, AI R&D cost, and total evolution time. It is designed to make model experimentation faster, more observable, and more reliable. - **Team:** [Aliaksei Kaliutau](https://cerebralvalley.ai/u/akaliutau) - **GitHub:** https://github.com/akaliutau/evolyth - **Demo video:** https://drive.google.com/file/d/1snxWgdQFZkDT1vJVehFpeSW7J9Kb0L3k/view - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=142 ### 41. Lekha-Jokha AI Lekha-Jokha is an AI-based investigation tool that aims to empower citizens, journalists, and scholars to make sense of the discrepancy between government claims and actual available public data. In India, important information related to politicians, public expenditure, audit reports, and various schemes exist. But there is no single source where all such information can be found. Hence, it becomes highly difficult for any citizen to hold their government accountable based on factual data. This system leverages the Claude Opus 4.7 agent as a reasoning engine to orchestrate multiple domain-specific agents (Assets, CAG, Schemes, News, and RTI). On receiving an input query in natural language, our system: Extracts structured information regarding politicians’ declaration of assets Finds out pertinent CAG audits reports Analyses allocations and utilization of schemes Provides news signals Finds RTI requests to bridge data gaps Then all this extracted information gets aggregated into a structured dossier of "Claimed vs Actual" discrepancies with proper citation. Unlike the conventional approach of dashboards, the Lekha Jokha is an AI-powered investigator that finds discrepancies in the fragmented data available in the public space. - **Team:** [Harshanand sharma](https://cerebralvalley.ai/u/Harshdev) - **GitHub:** https://github.com/Harshcoderhacker/lekha-jokha - **Demo video:** https://youtu.be/MI4h7Mc-3oI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=150 ### 42. Caremory Here's a tight submission description (~150 words): Caremory gives family caregivers in paper-heavy healthcare systems a single source of truth for their family's medical history — powered by Claude Opus 4.7. In MENA, South Asia, and most of the developing world, prescriptions and vaccine records live in drawers and camera rolls. When a child is sick at 2am, the paper system fails exactly when it matters most. Claude Opus 4.7 powers three features: photograph any prescription (Arabic, English, or mixed) and Claude extracts every medication with dose and frequency into a structured record; ask plain-language questions like "last antibiotic?" and get precise answers with tap-through links to the original document; generate a doctor-shareable link and a new physician sees an AI clinical summary plus full visual timeline — 18 months of history in 30 seconds, no login required. Multi-user access, current medications view, manual entry, and full PWA support. - **Team:** [Osama Al-Atroush](https://cerebralvalley.ai/u/osama-alatroush) - **GitHub:** https://github.com/OsamaAlatroush/Caremory-hackathon - **Demo video:** https://youtu.be/4h6UZUowRvM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=186 ### 43. Fathom Mode Most AI agents fail not because they reason poorly, but because they reason faithfully from a misread intent. Fathom Mode for Claude targets this failure mode, called Intent Decay: at 85% per-step accuracy, a 10-step agent pipeline delivers only 19.7% end-to-end reliability. It is a planning-session protocol that runs before Claude acts, in both Claude Code and Cowork. Opus 4.7 carries the dialogue: it restates its understanding, surfaces missing details, catches technical misconceptions, and asks one focused question per turn. A deterministic Python core handles the rest, maintaining an Intent Graph, the Fathom Score, Causal Provenance, and a zero-LLM compile step that turns the dialogue into a reviewable plan. The key invariant: Claude may infer supporting structure, but user-stated causal reasoning is kept separate from system-inferred structure before any code is written. After one approval, Claude executes from the compiled plan. The demo shows two real tasks: a daily-work review producing a personalized PDF, and a chaos-theory visualization where Fathom catches the user's Lotka-Volterra misconception in the first turn. - **Team:** [ziwei ma](https://cerebralvalley.ai/u/ma-ziwei) - **GitHub:** https://github.com/ma-ziwei/fathom-mode-for-claude - **Demo video:** https://www.youtube.com/watch?v=Nxz_zD05Id8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=25 ### 44. The Feed Looks Back The Feed Looks Back is a live performance built around the feeling of being almost, but not quite, seen: the texture of the contemporary feed made into a stage event. A musician plays. Claude Opus 4.7 listens once per phrase, reads the moment, and authors what should appear on screen: text that fades in, images that surface, live p5 sketches and SVG fragments, motion bound to amplitude and structural events in the music. Earnest, attentive, almost right. The gap between what the music is and what the screen returns is the piece. The default move when generative AI meets culturally specific material is to engineer the mistranslation away: fix it, hide it, retrain past it. I built a stage where it stays. The musician plays in a tradition outside what the model was optimized to recognize. Opus, given audio features, current scene state, and prose context about the musical world behind the performance, renders what should appear on screen with full attention. The result is sincere, sometimes beautiful, never quite at home: algorithmically personal, never quite for you. The system runs in two modes. Live Mode plays as the music plays. Python DSP extracts features, a Node runtime assembles each per-cycle prompt with the current scene state and recent decisions, Opus authors the next visual gesture, and the browser executes it. Presence, latency, and the visible risk of failing in front of an audience. Bake Mode is the offline composer. The same pipeline runs against a recorded track in three Opus passes. A composition pass takes the whole work as multimodal context, including spectrogram, DSP panels, per-cycle prose summaries, mood board, and reference photos, and writes a per-cycle visual score. An execution pass generates each cycle under that plan. A critique pass has Opus review its own work and rewrite the cycles that do not land. The refined score replays deterministically in sync with the audio. Composition, in the literal sense: thinking across time. What changes between modes is the temporal stance, Opus reaching live, or Opus composing across the whole piece. - **Team:** [Ammer Ayach](https://cerebralvalley.ai/u/amay01) - **GitHub:** https://github.com/amrayach/feed-looks-back-spike - **Demo video:** https://youtu.be/8pCh-nqXe2A - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=26 ### 45. Atnia StinKit is an open-source code intelligence platform that indexes your codebase into a structural dependency graph and lets you query it nine different ways — through a CLI, a web dashboard, or an AI agent via MCP. WHAT IT DOES (the three things nobody else has): 1. Architecture Diagram Comparison (stinkit see) — Upload a whiteboard photo, Lucidchart export, or any diagram. Opus 4.7's vision reads it at 3.75MP, extracts every service and connection, then compares it against your actual code graph. "Your diagram is 58% accurate. 2 phantom services. 3 missing connections. Here's the corrected version." No other tool on earth does this. 2. Codebase Q&A (stinkit ask) — "What would break if I removed the cache layer?" StinKit queries the structural graph for context, then Opus 4.7 explains the architecture using real file names and dependency counts. Not keyword search. Graph-augmented reasoning. 3. AI Refactoring Planner (stinkit plan) — "Replace sessions with JWT." StinKit analyzes every affected file, determines the safe change order (leaves before roots), and generates a sequenced migration plan with PR boundaries, effort estimates, and rollback points. PLUS: Blast radius check in <2 seconds (fully offline), real-time file monitoring (stinkit watch), professional audit reports (8-section, Deloitte-grade with health scores and remediation roadmaps), incident forensics (stinkit trace), and an MCP server with 9 tools so Claude Code and Cursor can query your code graph autonomously. THE NUMBERS: 167 source files. 407 tests passing. 17 API routes. 10 dashboard pages. 9 CLI commands. 9 MCP tools. 11 languages supported. MIT licensed. $0 to run. WHAT MAKES IT DIFFERENT: Every output shows what StinKit knows AND what it doesn't — "Local completeness: 75% · 956 ambiguous local calls · Known blind spots: event emitters, DI containers." Other tools claim 100% accuracy. We show you the real number. That radical honesty is the product. - **Team:** [Aritra Sarkhel](https://cerebralvalley.ai/u/Aritra) - **GitHub:** https://github.com/Aritra003/stinkit.git - **Demo video:** https://drive.google.com/file/d/1wtgEF3H_-W7mUY2UQKYZx23XrYSQ8wep/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=44 ### 46. Nina Kairos is a thinking layer for product builders. The bottleneck in shipping products has shifted from execution to judgment. Product builders are shipping faster than ever, but more speed doesn't automatically lead to better decisions. Most AI tools accelerate this problem rather than solve it - they're biased toward action, ignore unknowns and assume context instead of truly synthesising it. Kairos does the opposite. The product is structured around four stages. Signals is the data backbone: every piece of context a builder is exposed to - customer calls, research notes, tickets and internal threads, pulled into one layer, with sensitive material excluded by design. Two Claude Managed agents run weekly to ingest market and competitor updates automatically, so the system feeds itself between sessions. Synthesis is where judgment starts. Kairos looks at signals through three lenses simultaneously: internal (strategy, leadership, team debates), external (market and competitive moves), and customer (calls, tickets, direct feedback), and produces a single synthesised position. The value is in the connections across lenses: a customer signal that maps to a leadership-flagged risk, an internal capability gap that intersects with a market shift. Opportunities is where Kairos pushes back. Each ranked opportunity comes with a Devil's Advocate panel where the system argues against its own ranking. The opportunity layer also tracks which initiatives get picked and how they play out over time, so intelligence compounds: future synthesis flags when a builder is about to repeat a past mistake or miss the same kind of opportunity twice. Solution turns a chosen opportunity into three differentiated approaches with trade-offs named clearly, then pushes a structured brief to Linear: a project for the strategy and individual issues for the work, ready to hand to Claude Code for execution. I built Kairos because I urgently needed it myself. The bet underneath is that we don't just need to build faster, we need to build better, which means deciding what not to build, and when not to build it. A week in, I'm already making different decisions than I would have without it. - **Team:** [Nina Mannheimer](https://cerebralvalley.ai/u/ninamhmr) - **GitHub:** https://github.com/nina644/kairos - **Demo video:** https://www.loom.com/share/18ae85961bef4fe0adad1676fefc63a3 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=88 ### 47. Finalrun FinalRun is an AI-driven tool that completely reimagines mobile app testing by moving away from brittle code and toward vision-based intent. 1. Vision-Based Execution: Instead of looking for code-level IDs, it uses multimodal AI to literally "see" the app on an emulator or simulator just like a human user would. It understands what is on the screen and performs actions (tapping, swiping, typing) to complete tasks without relying on fragile selectors. 2. Plain English Specs: Tests are defined in readable, plain English YAML files rather than complex code. This ensures tests are easily understood by both human developers and AI agents. 3. Codebase Synchronization: The tool integrates directly into the AI coding environment. As code is written, it reads the source code and automatically generates test specs that live alongside the repository, eliminating the issue of "test drift." 4. The Autonomous "Verify & Fix" Loop: It integrates seamlessly into CI/CD pipelines. When a UI change breaks the app, it catches the failure, generates a report with device logs and a video, and feeds this data directly back to the AI coding agent. The agent then analyzes the logs, finds the root cause, and automatically fixes the code. - **Team:** [Ashish Yadav](https://cerebralvalley.ai/u/droidash) - **GitHub:** https://github.com/final-run/finalrun-agent - **Demo video:** https://www.youtube.com/watch?v=CRJs8SXKjRQ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=125 ### 48. Edtools Edtools Labs is an AI tutor powered by Claude Opus 4.7 that teaches STEAM through voice, live drawing on a shared canvas, and smartphone sensors—so any phone becomes a simple interactive lab. Goal: democratize rigorous, hands-on science learning globally, in a learner’s own language. - **Team:** [Andres Salcedo](https://cerebralvalley.ai/u/Soyandresalcedo) - **GitHub:** https://github.com/soyandresalcedo/Edtools-labs - **Demo video:** https://www.youtube.com/watch?v=3KAeiqsVQq8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=187 ### 49. Project Ellis Formcutter is an AI assistant that turns long, intimidating USCIS immigration forms into a guided conversation. Built for underserved communities where one DOJ-accredited nonprofit serves 13,400+ immigrants in NY's Southern Tier alone, it uses Claude Opus 4.7 to read user-uploaded documents (passports, tax returns, paystubs, IDs), extract sponsor and beneficiary information, and walk applicants through the remaining gaps in plain English across ten languages. Complex cases are routed to a real DOJ-accredited representative for review before filing — the AI handles the deterministic 80% of immigration paperwork; humans handle the 20% that needs judgment. We support the I-864, I-130, N-400, I-485, I-90, I-765, I-131, and I-589. - **Team:** [Aditya Kapoor](https://cerebralvalley.ai/u/adityuhkapoor) - **GitHub:** https://github.com/adityuhkapoor/formcutter - **Demo video:** https://youtu.be/K7qGsYDr61U - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=188 ### 50. Marvin Stevens Nagi is an AI elder-care product where the AI's job is to route routine requests and to step aside the moment a human family member should take over — and where that family member can self-host the entire stack on a $5 server they trust or own. The world is aging, and there are fewer younger family members available to provide care. It's a problem that grows every year, but one we can start preparing for now. Nagi is built on three capabilities that only recently became viable at consumer level: 1M-token context, prompt caching, and LLM-quality voice conversation — all wrapped in privacy and elder-paced interaction, so trust isn't an afterthought. We're building for what's next because the technology already crossed the threshold that makes this work, and most builders haven't noticed yet. The same architecture that lets one family run Nagi on a $5 server lets governments and NGOs digitize their pro-bono outreach at low cost . The goal is reducing isolation for the elders families struggle to reach; and for those with no family to reach them at all. - **Team:** [Steve R](https://cerebralvalley.ai/u/Stevens) - **GitHub:** https://github.com/SperLat/NaGi - **Demo video:** https://drive.google.com/file/d/1RkhCP1SHrcl4FUcU3AwRpWerOlyv87CX/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=196 ### 51. TeamAlma 6:30am, Lima. María, 23, is awake before the rest of the world — alone, after another hard day she didn't tell anyone about. Her phone buzzes once on the nightstand. The teal glow of the screen lights the wall. "Did you eat breakfast yet? ☀️" That's Alma. She wasn't asked. She wasn't summoned. She noticed. This is the entire idea behind Alma — a proactive AI emotional companion for people living with depression and loneliness in Latin America. The radical bet: most people who need help never reach out, so the AI has to come to them. ──────────────── WHAT ALMA DOES ──────────────── • Sends unprompted check-ins at 08:30, 13:30, and 19:30 Lima time via Telegram (APScheduler running inside the agent process — no separate scheduler service) • Remembers everything across conversations: a comment from 3 days ago can shape today's response (FastMCP server, SQLite + fastembed semantic embeddings) • Routes intelligently between two Claude models: claude-haiku-4-5 for everyday turns, claude-opus-4-7 the moment a deterministic crisis scorer flags distress (>0.7) • Receives images through Telegram and responds to them with full multimodal context — a photo of an empty plate captioned "finally ate" gets a response that sees both • Available simultaneously on Telegram and web chat (nginx → FastAPI), unified user identity ──────────────── SAFETY BY DESIGN ──────────────── Crisis detection is deterministic by choice. A keyword-based scorer (0–1) — not an LLM — gates every proactive message. Score > 0.6: proactivity suppressed. Score > 0.7: escalate to Opus 4.7. The safety layer cannot hallucinate. This is an architectural principle: deterministic safety, LLM response quality. Two separate concerns, two separate systems. ──────────────── MEMORY ARCHITECTURE ──────────────── A dedicated FastMCP server exposes 5 tools to Alma's chain: get_memory_tool, search_memories_tool, upsert_memory_tool, build_context_tool, evaluate_crisis_risk_tool. The crisis tool is dual-purpose — used by the conversation chain in real time AND by the APScheduler before each proactive send. Same logic, two contexts, never out of sync. ──────────────── TECHNICAL SHAPE ──────────────── 8 GitHub repositories, 5 Docker services, full Docker Compose orchestration: • claude-hackathon-agent — FastAPI + AlmaChain (LangChain) + APScheduler • claude-hackathon-mcp — FastMCP memory server • claude-hackathon-telegram — bot polling, stores chat_id in Redis for proactivity delivery • claude-hackathon-web — HTML/JS + nginx reverse proxy • claude-hackathon-infra — Docker Compose orchestrator Redis manages shared state across services: alma:proactive:crisis_score:{uid}, alma:proactive:last:{uid}, alma:proactive:slot:{uid}:{date}:{slot}. ──────────────── WHY THIS MATTERS ──────────────── 1.6 psychiatrists per 100,000 people. That's the number for Latin America. The number that matters most is invisible: the people who would benefit from mental health support but will never seek it. They don't show up in any statistic because they don't exist in any system. Alma exists for them. Every existing mental health tool — every chatbot, every meditation app, every therapy platform — assumes the user takes the first step. That assumption is the gap. People in acute depression do not take first steps. The act of opening an app and asking for help requires energy that depression specifically takes away. So Alma takes the first step instead. For someone alone at 6:30am in Lima who hasn't told anyone they're struggling, Alma sends: "Did you eat breakfast yet? ☀️" A small thing. But it means: I see you. You are not invisible. Someone reached out before you had to. ──────────────── WHAT MAKES THIS REAL ──────────────── That message takes 0.4 seconds to deliver. The architecture behind it took 12 specialized Claude Code agents, 2 rounds of design debate, 8 GitHub repositories, deterministic crisis gating, dual-model routing between haiku-4-5 and opus-4-7, persistent semantic memory across SQLite + fastembed, and a $200 personal commitment to keep Opus 4.7 in the loop exactly where it matters most. Mental health support that scales without diluting the response cannot be built with shortcuts. It has to be wrestled with. I wrestled with it. Alma is the result. This is not a prototype. It is deployed, proactive, listening, and reaching out — right now, three times a day, to anyone who connects via Telegram or web. The ratio of psychiatrists per 100,000 doesn't change overnight. But the number of people who hear from someone at 6:30am can. That's the bet. - **Team:** [Cristian Lazo](https://cerebralvalley.ai/u/Cris) - **GitHub:** https://github.com/iDeepBrain/claude-hackathon-alma - **Demo video:** https://youtu.be/9lSLd4g-0EM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=228 ### 52. Two+Aliens God Bless The United States Of Aliens is a non-looping, eighteen-minute AI-voiced ceremony for the 61st International Art Exhibition — La Biennale di Venezia 2026, where it appears as a guest work in Domus Diasporica (Pavilion of Bosnia and Herzegovina, curated by Isidora Živković, commissioned by Sarita Vujković, Palazzo Malipiero, May 9 – November 22, 2026). A visitor scans a QR at the pavilion, puts on headphones, taps Begin, and hears a private mass: four non-human voices — an EYEWITNESS filing a surveillance report, a CHURCH LEADER addressing the congregation, a CHORUS responding with the refrain, and a PRAYING ALIEN's interior monologue across three signals (brain-electric, heart-pulse, gut-rumble). A full ensemble carries the one-word AMEN and HUGS moments. The text — a short story I wrote on Christmas Eve 2024 about a protesting Christmas-Eve mass and overseas-based beings — was minted as the USALIEN ERC-721 on the Base blockchain that same day. The performance now lifts that static text into an active liturgical object. Claude Opus 4.7 is co-director. Each of the fifteen scenes is read by an Opus 4.7 structured-output call (Zod schema, adaptive thinking on, prompt caching) that casts the voices, sets per-segment pacing and reverb presets, and emits SSML for Kokoro-82M to perform. The director isn't a chatbot — it's a function: scene markdown in, validated direction JSON out. The piece operates across three frames of "alien" simultaneously — legal (the artist as non-EU non-citizen in Brussels), technological (AI as alien intelligence reading the mass from outside the species), cosmic (the story's surface narrative of overseas-based beings). The engine ships open-source from end to end — EUPL-1.2 for code, CC BY-NC-ND 4.0 for the artwork — where most AI art ships only its outputs. Made at Hectolitre, Brussels — the residency where the work was developed — and made possible by the $500 in Claude Code API credits, without which a Brussels-based, unbanked artist would not have been able to co-direct this with a frontier model. - **Team:** [Chris-Armel Iradukunda](https://cerebralvalley.ai/u/daqhris) - **GitHub:** https://github.com/daqhris/god-bless-usa - **Demo video:** https://daqhris.com/god-bless-usa/submission.html - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=229 ### 53. Kiln AI is quietly flattening everyone's voice. The more I used Claude and ChatGPT to write faster, the less I sounded like myself. Kiln is a Mac app that fine-tunes a local LLM on your own writing in 20 minutes. You drop a folder, or connect Notion and Apple Notes through MCP. An Opus 4.7 orchestrator spawns sub-agents in parallel to clean your corpus. Three classifiers we distilled from 5,000 Opus labels run locally to judge quality, preference, and style. Once trained, your voice runs on Ollama. And the cool part: Kiln also exposes it as an MCP server. So you can plug it into Claude.app and Claude will write as you. - **Team:** [Timothee Tavernier](https://cerebralvalley.ai/u/Timothim) - **GitHub:** https://github.com/timothim/kiln - **Demo video:** https://youtu.be/XFj-7J0CyQU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=231 ### 54. Alfred I operationalised my cognition into a tool for writers, solving the problem - how to turn high-dimensional thinking into legible artifacts without losing the author's voice. You spend 30% of your writing time just... **moving things around**. Not writing new ideas. Not researching. Just reorganizing. You have three paragraphs saying the same thing. You have one paragraph doing too much. You have ideas in the wrong section. You know what you want to say, but untangling it, so that others understand exactly what you mean, is a very time taking task. All AIs that try to solve this whitewash your voice. Meet Alfred, it is the Photoshop for writing structure. You write messy. It photoshops it into coherence. **It doesn't rewrite in AI voice. It refactors your structure while keeping your words.** Alfred constrains AI to operate only on structure — Split, Merge, Move, Hoist — and shows you in real time what it's learning about how you think. The thesis is that architectural constraint, not prompt instruction, is what preserves authorship when AI enters the loop. - **Team:** [Kavin Sood](https://cerebralvalley.ai/u/itskavins) - **GitHub:** https://github.com/kavinsood/alfred.git - **Demo video:** https://kavinsood.com/video - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=233 ### 55. UCL medic Every year 70000 junior doctors in the UK rotate onto new wards with unfamiliar software and zero onboarding. The medical knowledge is never the problem. Navigating the system is. ScrubIn is an AI companion that sees any clinical interface and guides doctors through it step by step, with an animated cursor pointing to exactly what to click, referenced against NICE guidelines and clinical protocols in real time. I am a 5th year UCL medical student who lived this problem. While bedbound recovering from a ruptured patellar tendon, I lost a research opportunity because I couldn’t navigate an ophthalmology software system without someone physically showing me where to click. When I returned to the wards I saw this similar situation happen to many junior doctors on different wards. ScrubIn works with any system - EPIC, EMIS, SystmOne, Cerner, NICE guidelines, ClinicalTrials.gov. If you can see it, ScrubIn can guide you through it. Built with Claude Opus 4.7 vision capabilities. Today it navigates, tomorrow it automates. - **Team:** [Arinze Okemuo](https://cerebralvalley.ai/u/Rinzonrepeat) - **GitHub:** https://github.com/okemuoab-byte/scrubin - **Demo video:** https://youtu.be/71caA-498rk?si=Uejwu5LdVpZVQ45A - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=251 ### 56. Hey Moosh Intently is Life Ops web app: a thin UI over Claude Managed Agents that handle the recurring operations of a life. Three flows ship end-to-end — **daily brief**, **daily review**, **weekly review**. Every morning, your brief is already there: what you’re working toward and what needs your attention today, before the noise starts. Every evening, a review captures what actually happened. Your journal goes deeper into patterns and blind spots you’d never catch from inside a single day. Day by day, your days feed your weeks. A year of that and you’re not where you drifted — you’re where you chose to go. The build answers "Build From What You Know." For years Muxin ran Life Ops via Claude Code on top of Markdown. It worked for one technical user. Intently turns that personal system into something a non-technical person can install. - **Team:** [Muxin Li](https://cerebralvalley.ai/u/heymoosh) - **GitHub:** https://github.com/heymoosh/intently - **Demo video:** https://www.youtube.com/watch?v=iK_0-889fkg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=258 ### 57. last_me RobbyMD is a doctor-steered diagnostic trace for clinical encounters. Most clinical AI tools produce a transcript, summary, or note. RobbyMD focuses on the missing layer behind the note: the reasoning path. It converts a live doctor-patient conversation into timestamped turns, structured clinical claims, correction edges, working differentials, discriminating questions, and provenance-backed SOAP drafts. The core problem is that physicians already reason during encounters, but the trace is usually lost. If a patient corrects a prior statement, if one answer changes the differential, or if a question rules down a branch, that reasoning path is often reconstructed from memory after the visit. RobbyMD preserves it as structured, inspectable state. The physician remains in control. RobbyMD does not diagnose, prescribe, order, triage, or recommend treatment. It surfaces relevant claims, tracks superseded or excluded evidence, updates deterministic differential projections, suggests next-best discriminating questions, and drafts SOAP only from active claims with provenance. The technical bet is simple: Claude should handle language understanding and phrasing, while the substrate handles auditability. Opus 4.7 extracts structured claims, phrases questions, and drafts note sentences. The deterministic substrate validates, stores, links, supersedes, ranks, and preserves the clinical trace. RobbyMD is a research prototype, not a medical device. The demo uses synthetic or benchmark data only. Its goal is to show what comes after medical scribes: not just better notes, but a physician-steered reasoning substrate behind the clinical note. - **Team:** [Harneet Bali](https://cerebralvalley.ai/u/laststan) - **GitHub:** https://github.com/harneet2512/RobbyMD - **Demo video:** https://youtu.be/_NC3XiGVaVM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=265 ### 58. Keskin Medkit is a virtual clinic that lets you make every mistake before the real patients come. It's a voice-first AI simulator for medical students and newly graduated doctors. You manage scientifically generated AI patients in real-time voice conversation — take the history, order labs, read imaging, diagnose, and prescribe. The patients are real agents that simulate real people, with different personalities. Some are anxious, some are evasive, some downplay their symptoms, some ramble. You have to manage the communication as much as the medical reasoning — exactly like a real consultation. After every session, Medkit grades your communication, history-taking, and clinical reasoning with scientific citations from the latest published guidelines (NICE, ESC, AHA, GINA, GOLD). The attending physician powered by Claude Opus 4.7 cannot fabricate guidance — its citations come from a curated registry, so the marking sheet quotes the same evidence a real examiner would cite. Trainees can run a dozen cases an hour and get a cited, structured debrief on every one. That feedback loop doesn't exist anywhere else. - **Placement:** 1st Place - **Team:** [Bedirhan Keskin](https://cerebralvalley.ai/u/bedirhan) - **GitHub:** https://github.com/bedriyan/medkit-app - **Demo video:** https://youtu.be/6bN6hnx-A2A - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=28 ### 59. Poligent Poligent, a policy simulation platform, is a multi-agent system that takes a government policy document and streams parallel reasoning from four demographically-grounded UK population archetypes using Claude Opus 4.7 with extended thinking. The platform ingests policy documents (starting with the 2010 UK VAT rise), uses a supervisor agent to produce personalised briefings anchored in each archetype's specific income and spend profile, then runs four archetype agents in parallel, each reasoning in first person about what the policy means to their actual household budget. A reporter agent synthesises the four reactions into a structured policy brief with distributional impact analysis and concrete recommendations. Results are validated against IFS published distributional findings. Extended thinking is central to output quality: the archetypes don't produce generic policy commentary, they do the arithmetic against their own persona profile and cite specific numbers. The React dashboard at ai-policy-simulation.vercel.app streams thinking tokens, UK neural-voice narration (edge-tts, one voice per archetype), and the final policy brief in real time, backed by a FastAPI service on Azure App Service. All runs persist as JSONL for replay without API calls. - **Team:** [Syed Ali Abbas](https://cerebralvalley.ai/u/SyedAliAbbas) - **GitHub:** https://github.com/syedaliabbas1/AI-Policy-Simulation - **Demo video:** https://www.youtube.com/watch?v=W2CksPASaXs - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=45 ### 60. Crosswalk Crosswalk reads your AI governance policy and validates it against EU AI Act, NIST AI RMF, FDA AI/ML SaMD, and ISO 42001 in one pass. You get drop-in policy language for every gap, not a remediation roadmap. The pipeline is a 3-step prompt chain. Sonnet 4.6 extracts every concrete commitment from the policy as structured data. Opus 4.7 maps each commitment to framework requirements with verbatim evidence citations — every covered finding cites a specific commitment ID and a substring of that commitment's evidence quote, so you can audit Crosswalk's reasoning by hovering, not by trusting. Opus 4.7 then drafts remediation language matched to the policy's voice — preserving modal verbs, structure, and register. Microsoft RAI gets dense legalistic prose with numbered RS4.X structure. A flatter corporate policy gets institutional voice. I empirically verified by running the same Article 55 gap through both and comparing. Two features I'm proud of: Industry-aware applicability gate. Tell Crosswalk your industry and role, and it marks structurally out-of-scope requirements as not-applicable. Run a tech company through FDA SaMD and clinical-device requirements get skipped with a written rationale per item, not surfaced as false-positive gaps. The skipped section is the difference between a tool and a checklist. Tier C cross-framework merging. Equivalent requirements across regulators (e.g., EU Article 9 risk management ≡ ISO 42001 Clause 6.1) collapse into single proposals via union-find clustering, so accepting one fix counts toward both regulators' coverage. Soft-conflict detection (Jaccard similarity on token sets) flags when regulators want different things on the same topic. Word .docx export ships with a cover page, framework citations, severity badges, change-list table, and full audit trail. Per-step prompt caching keyed on policy_hash + framework_id makes re-runs near-instant and deterministic. Schema robustness via permissive parse + post-hoc hydration handles model drift without failing. - **Team:** [Madina Gbotoe](https://cerebralvalley.ai/u/Zhalianna) - **GitHub:** https://github.com/mgbotoe/Crosswalk - **Demo video:** https://www.youtube.com/watch?v=UamV_92DOY0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=89 ### 61. Rohan Alexander Data cleaning is a huge pain. If it's done by a coder then it takes a huge amount of their time, but if it's done by someone who can't code then the analysis isn't reproducible and they have to redo their same steps the next time there is a data update. Sift bridges this gap by using Opus 4.7 to check for a bunch of data issues and then proposing changes which the non-technical analyst can approve/decline which removes any black-box concerns. At the end both the cleaned data and the underlying generated Python script can be downloaded so that the same cleaning steps can be run again when the data update and it's easy for a coder to add the changes into a process that can run automatically. - **Team:** [Rohan Alexander](https://cerebralvalley.ai/u/rohanalexander) - **GitHub:** https://github.com/RohanAlexander/sift - **Demo video:** https://youtu.be/BxveNHHHWlE - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=126 ### 62. HexalLabs Seven specialized AI models reason independently in parallel on a hex grid. Anonymous peer review removes groupthink — each model critiques others without knowing who said what. Apex (Claude Opus 4.7) synthesizes confidence scores and critique depth into one final answer that's demonstrably better than any single model. Includes Oracle mode (single model), The Relay (mid-chat handoff), Scout (web grounding), Primal Protocol (caveman rewrite) for the demo use hexallabs.com - **Team:** [Anas Abuelhaag](https://cerebralvalley.ai/u/anaselhaag) - **GitHub:** https://github.com/a-elhaag/hexallabs - **Demo video:** https://youtu.be/rK6DNoT6Wy4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=129 ### 63. Ernest Moyo (Solo) In 2025, 86% of USAID's 6,200 awards were terminated. Global Fund Replenishment 8 fell $5.36B short. The Lancet projects 9.4 million additional deaths by 2030. What collapsed wasn't drugs or supply chains — it was the labour layer: the proposal writers, M&E analysts, and indicator triangulators who turn raw data into fundable programmes. Jhpiego cut 1,975 staff across 44 countries. Bill Gates told Congress the Gates Foundation cannot backfill. SwarmAid replaces that layer with a swarm of 10 specialised agents — running on Opus 4.7, Sonnet 4.6, and Haiku 4.5 — orchestrated the way a real planning team works. EpiProfiler and LitReviewer run concurrently, pulling live data from WHO GHO, DHIS2, the Malaria Atlas Project, OpenAlex, and the Global Fund Data Service. SustainabilityAdvisor races four financing scenarios in parallel — Global Fund-only, bilateral donors, domestic resource mobilisation, and blended — and selects the strongest five-year trajectory. ProposalWriter drafts a 40-page modular concept note. Five adversarial personas — a TRP reviewer, NMCP officer, investigative journalist, LFA auditor, and biostatistician — attack the draft simultaneously. A Receipt Wall traces every numeric figure back to its source API call. The system is not locked to one country, donor, or disease. The demo runs on Malawi malaria with the Global Fund, but the architecture supports any country and any funding modality — including domestic health budget planning. Human-in-the-loop controls (Pause / Resume / Stop) are wired end-to-end, so a ministry official is always in control. Output: All things work outputs, plus Annex B (Receipt Wall), Annex C (Adversarial Report), and Annex D (Sustainability Trajectory). Generated in approximately 8 minutes. Open source, MIT licensed. - **Team:** [Ernest Moyo](https://cerebralvalley.ai/u/afrinestoe) - **GitHub:** https://github.com/ernestmoyo/swarmaid - **Demo video:** https://youtu.be/7JTinKZpxVc - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=130 ### 64. Citare One person processes a paper with AI; everyone should inherit the result. LLMs can read papers, but PDF reading is not claim infrastructure. PDFs are long, messy, and parsed inconsistently: one run may drop a mediator or boundary condition, while another turns “associated with” into “causes.” And the more important the paper, the more often this unstable process is repeated. For field-defining papers, it is plausible that a million people will ask AI to read the same paper over the next decade. That wastes energy and amplifies citation drift. Citare uses Claude Opus 4.7 to convert papers into reusable claim graphs. Each claim stores source text, page numbers, evidence type, causal strength, safe citation verbs, boundary conditions, and integrity warnings. Other tools return metadata; Citare returns the actual claim and makes it reusable through an open database and CitareMCP. Try it now at https://citare.dev/ - **Team:** [Ryosuke Ishii](https://cerebralvalley.ai/u/ryouen) - **GitHub:** https://github.com/ryouen/Citare - **Demo video:** https://youtu.be/syp6Qxab6xQ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=189 ### 65. AAR TaluGPT is a natural-language map of Estonian local food. 2,130 farms, producers, markets, shops and food events from the PTA organic registry, e-Äriregister, avatudtalud, kohaliktoit, EPKK and laadakalender, unified into a single searchable Leaflet map. Filter by county, food category (milk, eggs, honey, etc) or organic-only, then ask the embedded AI assistant Claude Opus 4.7 + Qdrant + n8n, questions like "Where to buy organic honey in Tartu?" and get an answer with the relevant farms highlighted on the map. Estonian and English in one space, but data based of Estonian farms, events etc. Live at https://talugpt.vercel.app - **Team:** [Arle Aron Reidma](https://cerebralvalley.ai/u/AAR1) - **GitHub:** https://github.com/AARenor/talugpt - **Demo video:** https://youtu.be/UrSwSuQJu5o - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=27 ### 66. Synapse — The Quality Gate Synapse is an AI quality gate that measures and optimizes prompt inputs before they reach a model, and prevents bad inputs from being executed. Built on Anthropic's Opus 4.7 Managed Agents API, Synapse uses a custom tool (PQS, the Prompt Quality Score) to grade every input across eight dimensions: clarity, specificity, context, constraints, output format, role definition, examples, and chain-of-thought structure. When a prompt grades D or F, the agent refuses to answer it and rewrites it instead, then re-scores the rewrite to verify the lift. The same gate runs in two surfaces. The human-facing demo at synapse.promptqualityscore.com lets anyone paste a prompt and watch Opus 4.7 grade, refuse, rewrite, and re-grade in real time. The agentic-economy surface at /marketplace shows two autonomous agents transacting on-chain via x402 USDC on Base mainnet, with the same quality gate running between them. AI input quality is the missing measurement layer in the AI economy. Conservatively, ten billion dollars in compute is burned annually on inputs that were never going to work. Synapse is the gate that reclaims it. - **Team:** [Ken Burbary](https://cerebralvalley.ai/u/pqs) - **GitHub:** https://github.com/OnChainAIIntel/synapse-demo - **Demo video:** https://youtu.be/sbeiYIjw-uo - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=46 ### 67. Finerium NERIUM is infrastructure for the AI agent economy. Five integrated pillars: Marketplace, Builder, Banking, Registry, Protocol. Packaged as a playable JRPG world to make agent infrastructure accessible to non-technical founders. The Builder pillar is a gamified prompting editor. A user types one sentence describing their project. Sekuri, the agent structure architect, classifies project complexity (small=4 agents, medium=8, large=14) and proposes a multi-vendor agent roster. Eight vendors are presented in the model selection interface: Anthropic Opus 4.7, Google Gemini Pro, OpenAI Codex, Higgsfield, Seedance, Meta Llama 3.1, Mistral Mixtral, and Auto. The agents spawn in parallel terminals and execute the build end to end. The Marketplace pillar solves the indie creator monetization gap. Today, AI skills sit forgotten on Twitter feeds, MCP servers buried in GitHub repos, brilliant prompts traded in Discord and lost by Monday. NERIUM Marketplace gives every creator a permanent home and every buyer a real search engine. Discoverable, rated, sortable by trust score, monetized from day one. The Banking pillar treats AI agents the way utilities treat electricity. Every query metered. Every execution billed transparently. Buyers pay only for what they use. Creators earn revenue per execution, automatically. Stripe Connect Express integration with usage-based metering. The Registry pillar provides cryptographic identity for every agent on the network. Verified, audited, and signed by the protocol itself. Ed25519 signing with weekly key rotation. Astraea Bayesian trust score with Wilson confidence interval recomputed nightly. The Protocol pillar provides multi-vendor adapter dispatch with automatic fallback. One agent can route across Claude, Gemini, custom models. Vendor neutrality built in. Circuit breaker and kill switch architecture for graceful failure. This submission was constructed by 54 specialist Claude Code agents on Opus 4.7, orchestrated via formal multi-version handoff documents (V1 through V7 in _meta/orchestration_log/). Each agent has a defined role, dependency contract, ferry discipline, and self-check verification. The recursive automation thesis: NERIUM was built by managing 54 Claude agents through manual orchestration, and NERIUM's Builder pillar gives that capability to its users. Honest claims: Stripe in test mode at submission, Stripe Atlas activation pending. Builder live runtime is theatrical at submission with bring-your-own-key pattern available for judges with their own Anthropic API key. Multi-vendor model selection UI is showcased; live runtime invocation at submission is Anthropic-only via Max plan workflow, multi-vendor live runtime activates per-vendor billing setup post-launch. Voiceover generated via Gemini 2.5 with bundled cyberpunk synthwave background music for consistency, demonstrates further multi-vendor AI usage acceptable per hackathon rules. Built by Ghaisan Khoirul Badruzaman, first-year Politeknik Negeri Bandung Teknik Informatika student in Indonesia. - **Team:** [Ghaisan K](https://cerebralvalley.ai/u/Finerium) - **GitHub:** https://github.com/Finerium/nerium - **Demo video:** https://youtu.be/DJQXitRa1VE - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=90 ### 68. Greatest of All Times PitOS is an AI-native team workspace for FIRST Robotics Competition teams. It replaces the chaos of group chats and spreadsheets with Slack-style channels, a kanban task board, a decision log, and a season recap generator — all wired into Claude agents that do real work rather than just chat. Key features: - @pitos in any channel triggers a streaming Claude reply with web search scoped to FRC-authoritative sources (FIRST, The Blue Alliance, Statbotics) - Judge simulator grounded in official FIRST Impact Award judging criteria — students practice before regionals with adaptive difficulty and a debrief - Persistent agent job queue processes background tasks (task extraction, memory updates, exit interviews) without blocking the UI - Full PWA — installable on iOS, Android, and desktop with offline shell and iOS splash screens - Mobile-first responsive layout with off-canvas sidebar and drawer panels Stack: Next.js 16 App Router · React 19 · Tailwind v4 · Drizzle ORM + LibSQL/Turso · Lucia v3 magic-link auth · SSE streaming · @anthropic-ai/sdk with claude-opus-4-7 Live demo: https://pitos.8092.tr - **Team:** [Bugra Canata](https://cerebralvalley.ai/u/canata) - **GitHub:** https://github.com/bcanata/pitos - **Demo video:** https://youtu.be/vj5YHWDRRVM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=93 ### 69. Mncedisi Bhembe Gas Wiser is an AI-powered fuel intelligence platform for US and Canadian drivers. Users input their vehicle from a real EPA fuel economy database, set home and work locations, and get auto-generated daily route costs. The Route Planner takes any origin and destination and produces a full AI analysis, fuel cost, optimal fill-up stop, efficiency score, price outlook, and vehicle-specific driving tips. The Trip Planner handles multi-day road trips with per-leg cost breakdowns, fill-up strategy, and departure timing. Everything is form-driven, no chat interfaces. Once a plan is generated, a single "Chat About This" button opens Claude with full context of that plan for follow-up questions. Powered by Claude Sonnet 4.7. - **Team:** [Mncedisi Bhembe](https://cerebralvalley.ai/u/rapthar) - **GitHub:** https://github.com/rapthar/gaswiser.ai - **Demo video:** https://youtu.be/fI4JktS9fNE - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=127 ### 70. ID2R - Ideas 2 Reality My project is a free tool that enables SMEs around south east asia to explore opportunities with AI by talking to a team of consultants that are ai agents who will help this business owner really find ways to implement and incorporate AI that will possibly improve their business - **Team:** [Paolo Gonzales](https://cerebralvalley.ai/u/aop) - **GitHub:** https://github.com/capturingthe3rd/fixmyworkflow - **Demo video:** https://youtu.be/v9iY-FtPVjY - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=190 ### 71. Godswill Lazarus Protocol is a crypto continuity insurance layer built for the $140 billion problem nobody talks about. Every year billions in crypto become permanently unrecoverable — not stolen, not lost to bad trades, just locked inside wallets that nobody alive knows how to open. We built Lazarus because we watched it happen to real people. Friends passing away. Families locked out. Money that could have changed lives just sitting there forever unreachable. Lazarus wraps around the wallet you already use. You never move your coins. You never give up your keys. You simply build a trust network — Confirmers who verify something real happened, Custodians who receive your assets, and Watchers who can veto anything suspicious. Your on-chain activity acts as a heartbeat. As long as you’re active nothing happens. The moment silence grows beyond your chosen threshold your network activates. The anti-collusion mechanic is enforced at contract level — the person who votes to confirm a release can never receive funds from that petition. Not a policy. Not a promise. Code. And the owner can always Resurrect with one signed transaction cancelling everything instantly. No new wallet. No custody. No fees while you’re alive. Ever. - **Team:** [Godswill SOL SOL](https://cerebralvalley.ai/u/SOL1M) - **GitHub:** https://github.com/DEVSOL404/lazarus-protocol - **Demo video:** https://youtu.be/R5TaMXyXpp4?si=rhy2oEHrNJDk7c-H - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=230 ### 72. Quixoticals ~$8B/yr in U.S. security deposits are wrongfully withheld. In England, 30% of TDS arbitrations land partly or fully for the tenant. In Pakistan there is no deposit-protection scheme at all - tenants in Karachi/Lahore routinely lose 1–3 months of advance to bogus deductions. The pain is universal: tenants don’t know the law, landlords do. DepositBack closes that asymmetry. The user uploads three things: their lease, the landlord’s deduction letter, and a few move-out photos. A single autonomous Claude Opus 4.7 agent then runs a multi-tool session that: Lists every rule in the active jurisdiction’s statute corpus (statute_lookup). Extracts text and page images from each PDF (pdf_read). Classifies each visible mark in every photo as “ordinary wear” vs. “chargeable damage” using Opus 4.7’s high-resolution vision (vision_analyze), citing the Connell v. Brownstein-Louis Co. (1927) standard for California and the equivalent rule for each other jurisdiction. Maps every line item in the deduction letter against statute, classifying each as valid, invalid, or excessive. Drafts a formal demand letter with verbatim statute citations (doc_draft). Pre-fills the appropriate court / scheme form for the jurisdiction (form_fill). The user watches the entire event stream live — every thinking step, every tool call, every result — over Server-Sent Events, then downloads the letter and the form. Adding a new jurisdiction is one JSON file in backend/data/statutes/. Five jurisdictions are live today: California (post-AB 12), New York, Texas, England & Wales, and Sindh, Pakistan (with Cantonments-Act fallback for DHA / Clifton Cantt / Faisal Cantt properties). DepositBack uses no RAG, no embeddings, no chunking — the entire lease, the entire deduction letter, the entire statute corpus, plus all photo descriptions fit in Opus 4.7’s 1M-token context in a single pass. This ~$8B/yr in U.S. security deposits are wrongfully withheld. In England, 30% of TDS arbitrations land partly or fully for the tenant. In Pakistan there is no deposit-protection scheme at all — tenants in Karachi/Lahore routinely lose 1–3 months of advance to bogus deductions. The pain is universal: tenants don’t know the law, landlords do. DepositBack closes that asymmetry. The user uploads three things: their lease, the landlord’s deduction letter, and a few move-out photos. A single autonomous Claude Opus 4.7 agent then runs a multi-tool session that: Lists every rule in the active jurisdiction’s statute corpus (statute_lookup). Extracts text and page images from each PDF (pdf_read). Classifies each visible mark in every photo as “ordinary wear” vs. “chargeable damage” using Opus 4.7’s high-resolution vision (vision_analyze), citing the Connell v. Brownstein-Louis Co. (1927) standard for California and the equivalent rule for each other jurisdiction. Maps every line item in the deduction letter against statute, classifying each as valid, invalid, or excessive. Drafts a formal demand letter with verbatim statute citations (doc_draft). Pre-fills the appropriate court / scheme form for the jurisdiction (form_fill). The user watches the entire event stream live — every thinking step, every tool call, every result — over Server-Sent Events, then downloads the letter and the form. Adding a new jurisdiction is one JSON file in backend/data/statutes/. Five jurisdictions are live today: California (post-AB 12), New York, Texas, England & Wales, and Sindh, Pakistan (with Cantonments-Act fallback for DHA / Clifton Cantt / Faisal Cantt properties). DepositBack uses no RAG, no embeddings, no chunking — the entire lease, the entire deduction letter, the entire statute corpus, plus all photo descriptions fit in Opus 4.7’s 1M-token context in a single pass. This eliminates the retrieval-error class of bugs that sank DoNotPay.eliminates the retrieval-error class of bugs that sank DoNotPay. - **Team:** [Muhammad Mufeez](https://cerebralvalley.ai/u/mufeezhanif) - **GitHub:** https://github.com/mufeezhanif/depositback - **Demo video:** https://youtu.be/9VOV1C75E9A - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=235 ### 73. Sightline Sightline is an out-of-home (OOH) advertising analytics platform with a 3D map view powered by Deck.gl and Mapbox. It helps advertisers and media planners visualize billboard locations, model pedestrian traffic, and generate AI-driven creative briefs — turning fragmented street-level data into actionable campaign insights. - **Team:** [Jason Suhari](https://cerebralvalley.ai/u/jasonms) - **GitHub:** https://github.com/jasonmatthewsuhari/sightline - **Demo video:** https://youtu.be/OUpO177KY-A - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=237 ### 74. Defensor Defensor is a Spanish-language AI legal advocate for patients in Latin America who have been wrongfully denied healthcare. A patient photographs their denial document with their phone. Defensor's Vision Agent reads it using Opus 4.7's 3.75 MP vision, extracting 15 structured fields (patient name, hospital, reason for denial, specialty requested, dates). The Violation Agent then loads Peru's full patient-rights law (Ley 29414) into the 1M context window and identifies exactly which articles were violated — no chunking, no retrieval, the whole law in one shot. The Channel Agent deterministically routes the complaint to the correct authority (EsSalud Defensoría, SUSALUD, or Defensoría del Pueblo) based on auditable config files. The Drafter Agent writes a formal complaint letter in legal-register Spanish, complete with real article citations and the statutory disclaimer on every output. Finally, a Follow-Up Managed Agent runs autonomously for 25 days — sending confirmations, reminders, and escalations at days 1, 7, 15, 20, and 25 — so patients are never left wondering what happened after they filed. Peru's Ombudsman receives ~4,000 healthcare complaints per year. Most are handwritten, poorly formed, or abandoned. Defensor gives patients with no lawyer budget a free, instant, precise legal advocate in their own language. It also works for Colombia and Mexico via swappable country-module configs. Tech: Next.js 16 + TypeScript (frontend) · Python 3.11 + FastAPI + Claude Agent SDK (6 agents, 21/21 tests passing) · Deployed: Vercel + Railway. - **Team:** [abel mancilla](https://cerebralvalley.ai/u/Arkhangio) - **GitHub:** https://github.com/arkhangio10/defensor - **Demo video:** https://youtu.be/6FZGwPGCARM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=252 ### 75. Handoff Handoff turns every disconnected hospital system into a single connected space where physicians can see what's at risk and build the tools they need. Physicians spend half their day at a computer, reassembling each patient's story from a dozen disconnected systems. The data exists. It just isn't accessible in any way that matches how a physician actually thinks about a patient. Handoff centralizes EMRs, staffing, insurance, and more into one live picture. Underneath, eight Claude Opus 4.7 sub-agents, each with its own set of skill files, read their domain and post observations onto a bitemporal knowledge graph. A Chain Finder reasons across them to surface risks. For longer questions, a Claude Managed Agent runs, producing persona-aware "what if?" verdicts grounded in the same data layer. On top of all this sits a widget layer that lets physicians shape their own view. Instead of being stuck with whatever dashboard a vendor decided on, they can spin up small, focused widgets for the patients, units, or risks they actually care about, and rearrange them into a workspace that matches how they think. The data layer feeds every widget, so anything new they build is live the moment it appears. Physicians have always worked inside the tools they were given. Handoff lets them build the ones they actually need. Check it out: handoff-claude.com - **Team:** [Noor Abdalla](https://cerebralvalley.ai/u/noorabdalla) - **GitHub:** https://github.com/noorabdalla04/Handoff - **Demo video:** https://youtu.be/2H4Q2maZA_8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=254 ### 76. Colored Bits coloredbits is a TypeScript library that gives every string in an LLM agent a provenance tag: every byte remembers where it came from — user input, tool output, retrieved document, model output, system prompt — and the runtime refuses to forget. Once a span is marked untrusted, the type system and runtime checks make it impossible to silently feed it into a higher-trust sink (a tool call, a downstream prompt, a SQL query) without an explicit, auditable downgrade. The hackathon submission ships a four-recipe cookbook that maps directly onto the prompt-injection literature: Chat rejects injection — refuse to dispatch a tool when the input would raise the conversation's trust floor. Spotlighting (Hines et al. 2024) — wrap untrusted spans in delimiters before the model reads them. Handle substitution (PFI / FIDES 2025) — seal high-trust spans behind opaque <> tokens so the model can reference them without ever seeing them. RAG with provenance — propagateLiteralFragments re-attaches source URLs to the substrings the model copied verbatim, recovering per-character citations automatically. The motivation is personal. I rely on agents every day as a developer, a founder, and someone with ADHD — they genuinely changed my life, and they are also one prompt-injection away from leaking my partner's calendar or rewriting a repo. The future is agentic, and it has to be responsible. coloredbits is the missing primitive: one axis the agent cannot lose ;) - **Team:** [Alex Vidal](https://cerebralvalley.ai/u/doublethinker) - **GitHub:** https://github.com/double-thinker/coloredbits - **Demo video:** https://youtu.be/LzNzPo6zTm8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=261 ### 77. Code by Hue Crescendo is an AI-powered project management tool where every AI response is backed by the complete state of your project — not just what you typed. The core innovation is the Board-State Awareness Engine (BSAE): a structured snapshot of your board — team workloads, member skill profiles, blocked tasks, sprint health, overdue items, and dependency chains — injected into every Claude Opus 4.7 call. This is what makes Crescendo different from Linear AI, Notion AI, and every other project tool with a chat widget: they see a card in isolation. Crescendo sees your whole project. When you ask "who should I assign this to?" Crescendo checks who has capacity AND whose strengths match the task. When a member is at risk of burnout, the AI refuses to assign without an explicit Overdrive acknowledgment. When you highlight text and click "Ask Crescendo," the suggestions are specific to your board state, not generic text transformations. Built with Claude Opus 4.7 for all AI reasoning, Claude Managed Agents for long-running backlog analysis, Claude Code for development, and fully open source under MIT license. - **Team:** [Kweku-Abeiku Attah Anyen](https://cerebralvalley.ai/u/kaanyen) - **GitHub:** https://github.com/Crescendo-PM/crescendo - **Demo video:** https://drive.google.com/drive/folders/12neSeOrFhNVdvW7lOxJnfJRxPR2zK7Rw?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=234 ### 78. Skim Intelligence Prediction markets pay bots to exist. Maker rebates, liquidity rewards, and mint/burn arbitrage are structural edges that pay regardless of which way a market resolves. The catch: surfacing those edges takes quant analysis most operators can't afford, so most leave the yield on the table. Skim Intelligence closes that gap with reasoning, not rules. A five-agent pipeline — three of them powered by Claude Opus 4.7 — autonomously scans markets across Polymarket and Bayse, reasons about each one across three strategy layers (mint/burn arbitrage, market making, reward farming), and paper-trades the structurally direction-neutral edges. The Alpha Agent streams its tool-use JSON token-by-token to a live dashboard, so you watch Opus 4.7 walk through the orderbook in real time before it commits to a recommendation. The Risk Agent runs an independent prompt as a circuit breaker. Execution simulates fills with realistic slippage and hard negative-EV guards. Every five minutes, Reporter writes an honest P&L attribution. Built on the Anthropic Messages API (tool use + streaming + ephemeral prompt caching), Cloudflare Workers + Durable Objects + D1, and a small Node relay for venue data. Fully open source under MIT — agent prompts, paper-trading engine, and orchestration all in the public repo. - **Team:** [Ojay Shaq](https://cerebralvalley.ai/u/ojay) - **GitHub:** https://github.com/ojaydev/skim-intelligence - **Demo video:** https://youtu.be/YaIPCAKgAr4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=47 ### 79. Dokely Deployment is complete. You can access Dokely directly at dokely.com. It is common for patients to feel lost when they are sick. Many do not know how to describe their pain, which specialist they should see, or what information actually matters before a visit. From the provider’s side, medicine has become highly specialized, so when the wrong case lands on the wrong desk, time is lost during transfers and the risk of misdiagnosis increases. Dokely is my attempt to bridge that gap between “patient language” and “doctor language.” I used Claude to help organize large scale U.S. public medical data and generate structured, evidence based physician profiles. The system works over public records for 156,113 California physicians, using sources such as NPPES, the Medical Board of California, OpenAlex, and CMS Medicare data. A patient can describe symptoms in everyday language, and Dokely converts that input into medical terminology, including MeSH based concepts, then matches it against academic papers, public physician records, and clinical billing patterns. The goal is not to diagnose the patient or recommend a doctor as if the system knows everything. Dokely is designed as a discovery tool, not a referral or booking service. It helps patients understand which physicians may be relevant to their situation and why, using transparent public evidence. It also includes red flag triage before any doctor card is shown, so emergency signals such as chest pain, stroke symptoms, or pediatric respiratory distress are handled as safety cases rather than ordinary search queries. Because Dokely does not hold private insurance contract data, it does not pretend to know whether a doctor is in network. Instead, it generates an insurance call script that patients can use to verify coverage, new patient availability, and cost estimates directly with insurers or clinics. This design choice matters to me: the system should be useful, but it should also be honest about the limits of its data. By connecting patient descriptions, medical terminology, academic evidence, and public physician records, Dokely creates a foundation that helps patients find a better starting point when it matters most. - **Team:** [jeageon lee](https://cerebralvalley.ai/u/autombel) - **GitHub:** https://github.com/jeageon/dokely - **Demo video:** https://www.youtube.com/watch?v=uCVYjViHAy4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=91 ### 80. Malte Landgren Orate is a library for Programmatic Grammar over Inference of LLMs. It unifies structured output, toolcalling under one abstraction, @program, which headline feature is that it allows you to constrain the fields you extract with arbitrary python predicate. Instead of only typing, now we can also guide the output with logic. The library handles compiling your decorated program schema into a grammar, and how it should recompile when needed depending on how the execution goes (as fields can reference other fields!). These programs are also composable, and it then also handles switching grammars. Today only rudimentary optimizations are in place, like grammar caching so we don't always need to recompute. - **Team:** [Malte Landgren](https://cerebralvalley.ai/u/Malte) - **GitHub:** https://github.com/maltelandgren/orate - **Demo video:** https://www.youtube.com/watch?v=SutZZU9Spds - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=128 ### 81. Rodrigo Passarelli kinn is a real-time Bayesian diagnostic interview engine for consultants, founders, and anyone running discovery calls. After every stakeholder answer, Claude Opus 4.7 computes the Expected Information Gain of every candidate next question against a Bayesian belief over the stakeholder's actual situation — and asks the one that collapses the most uncertainty. The frame is BED-LLM (Bayesian Experimental Design with LLM samplers): Opus 4.7 is *both* the answer-distribution sampler (predicting how a stakeholder might respond) and the belief updater (revising priors after each real answer). DSPy/GEPA compiles the prompts offline; runtime is a forced-tool-call loop with prompt caching, two-phase recompile, and dual-algedonic state separation. Built solo across the hackathon week in three iterations (kinn → kinn2 → kinn3). The dual-gate benchmark publishes the honest result (0.852 vs 0.920 target) rather than ship a hidden failure. - **Team:** [Rodrigo Passarelli](https://cerebralvalley.ai/u/Quidid) - **GitHub:** https://github.com/rpassarelli/kinn - **Demo video:** https://youtu.be/oIVqGh6W79w - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=132 ### 82. kamikos Kosa is an agentic Attack/Defense CTF toolkit. In a typical eight-hour ad ctf, a team has to repeat the same loop every minute: watch scoreboard, inspect suspicious network flows, reconstruct working exploits, patch vulnerable services without breaking the checker, and attack dozens of other teams before they fix the bug. At that pace, the challenge becomes less about individual hacking skill and more about keeping the whole workflow synchronized. Kosa wraps that loop around a Claude Opus 4.7 dispatcher. It ingests scoreboards from FAUST, Enoflag, and ATKLAB; consumes traffic from Tulip, pcaps, and JSON logs; reconstructs exploit candidates from observed flag-stealing flows; proposes patches; validates them against the live checker; and deploys through gated approval. Every action is auditable and constrained by explicit allowlists. Kosa supports two modes: manual mode, where Kosa acts as a dashboard and gated submitter, and autonomous mode, where the dispatcher drives the loop end to end. It is designed to complement existing A/D tooling such as Tulip for traffic capture and Ataka for exploit farming, rather than replacing them. Kosa also ships with a fully Dockerized three-team A/D lab, letting you run the entire workflow locally in about five minutes. - **Team:** [Kamil Orzechowski](https://cerebralvalley.ai/u/kamikos) - **GitHub:** https://github.com/kamikos/Kosa - **Demo video:** https://youtu.be/3svCBbPh61c - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=145 ### 83. 4ailabs Inferentia is a clinical AI system that models chronic disease as rigid predictive priors — not just abnormal biomarkers. Three coordinated Opus 4.7 agents run closed-form math across 22 metabolic nodes, select modulators from a nutrigenomic knowledge base, and synthesize a full clinical protocol with extended thinking. The same posterior produces two outputs: a clinician view with auditable math, and a patient view with a salutogenic narrative and body map. Built by a practicing integrative physician with two decades of clinical experience. - **Team:** [Miguel Ojeda](https://cerebralvalley.ai/u/4ailabs) - **GitHub:** https://github.com/4ailabs/inferentia - **Demo video:** https://youtu.be/war5LomovBk - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=146 ### 84. Kamino Corporation Corellia is a vendor-neutral control plane for AI agents — a single platform to spawn, deploy, govern, and manage agents at scale across any model, provider, or agent framework. The problem: As organisations begin deploying AI agents across their workforce, they hit a fragmentation wall fast. Deployment, tool permissions, observability, and orchestration are each solved by different point tools. None talk to each other, and most lock you into a specific stack. The ones that do always result in model and/or model vendor lock-in (e.g. AWS, Azure etc.) What Corellia does: Admins define the guardrails — which tools an agent can access, what model it runs, where it deploys — and Corellia enforces them. Think AWS IAM, but for agents. The v1 flow: Pick a harness from the catalog, give your agent a name and model, equip it with tools and scopes, configure its deployment, and launch — one Fly.io machine per agent, secrets-isolated, auto stop/start, pinned to an immutable Docker image digest so governance is auditable and reproducible. Under the hood: Go + Connect-go backend, Next.js 15 frontend, Proto IDL as the single FE↔BE contract, sqlc for typed SQL, Supabase auth with offline JWKS validation. The harness interface contract — how Corellia talks to any agent framework — is designed like the LSP: define the spec once, every new framework becomes a one-time integration. v1 ships with: - Hermes Agent (Nous Research) as the first harness - Tool-scope governance (URL / command / path allowlists per toolset) - Live chat panel per agent - Fleet management with bulk config operations Future adds: - Memory integration (from any provider incl. custom) - Skills registry - Deployment to more baremetal providers (AWS, Azure, GCP, Hetzner etc.) - Addition of more agent harnesses (e.g. OpenClaw etc) See it here: https://corellia-frontend.vercel.app/sign-in Login details: test@user.com + password: Test1234! - **Team:** [Philipp Holke](https://cerebralvalley.ai/u/pgbouncer) - **GitHub:** https://github.com/hejijunhao/corellia - **Demo video:** https://youtu.be/Xtx1NnX4vCA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=148 ### 85. Brandon Dent, MD Speak into the mic at https://prism42-app.thegoatnote.com/prism42/livekit and you reach a real-time 911 voice dispatcher that responds in under 1.5 seconds and walks a full emergency protocol: address verification, complaint-specific triage, and T-CPR pre-arrival instructions. This project began by stress-testing Claude Opus 4.7 on HealthBench Hard and MedAgentBench, then building a custom evaluation harness to exceed native model behavior—turning a general-purpose LLM into a safety-constrained, real-time clinical decision system. Prism42 developed out of a need for 911 call centers to have AI assistance during downtime. A deterministic finite-state machine owns every safety-critical decision and template — CPR instructions, repositioning, breathing verification — delivered verbatim. A parallel Opus 4.7 critic scores intent agreement across a 100-fixture evaluation. Naive LLM dispatchers drift and improvise; Prism42 supervises with an LLM and dispatches with an FSM. A cloud demo (https://prism42-console.vercel.app/prism42-v3 ) uses ElevenLabs + Claude for smooth playback; the full system runs self-hosted on a Brev B300 (https://prism42-app.thegoatnote.com/prism42/livekit ). Every life-safety path is physician-reviewed and traceable to NHTSA EMD, AHA T-CPR, and NHS Pathways. 83/83 tests pass. MIT licensed. When most AI systems go silent, Prism42 keeps responding. - **Team:** [Brandon Dent](https://cerebralvalley.ai/u/GOATnote) - **GitHub:** https://github.com/GOATnote-Inc/prism42 - **Demo video:** https://youtu.be/VIGRD6-1k8c - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=191 ### 86. TheCreator SapientIA — Inclusive Socratic Tutor The problem: Generative AI made answers free and learning rare. Students paste, they don't reason. The learners who were already underserved — blind, dyslexic, ADHD, motor-impaired — get hit twice: first by a tool that won't teach, and second by an interface that wasn't built for them. What we built: A Socratic tutor that refuses to answer turn 1. It diagnoses, asks leading questions, and gives graded hints (hint_1 → hint_3). Final answers only land after the learner demonstrates understanding. Every concept earned through reasoning is tagged earned; every concept the tutor had to hand over is tagged told. The earned ratio is the headline metric — the first tutor that measures the difference between learning and Googling. Why it's different: Accessibility is woven into the model's behavior, not painted on the UI. The learner's profile flows into both the system prompt (reading register, sentence chunking, single-question rule, define-on-first-use) and the interface (Atkinson Hyperlegible for dyslexia, ADHD focus-fade, dual ARIA-live regions, focus rings). Four axes — Visual · Cognitive · Learning · Motor — that compose orthogonally: a blind, dyslexic learner with ADHD can stack every layer and still get the same Socratic teaching, with no "lite mode." Hands off the keyboard. Hold Shift+V, say "recap", "slow down", "minimize". Nine commands, no menus, no setup. Voice activation cancels in-flight TTS (barge-in) so a learner who can't reach a keyboard never gets stuck waiting. - **Team:** [Bryan Vela](https://cerebralvalley.ai/u/Branelio98) - **GitHub:** https://github.com/bryanvela98/SapientIA - **Demo video:** https://youtu.be/r_oirNbrDsw, https://drive.google.com/drive/folders/12VfhHQZAKb79s8nVE57rG3gAOYrpPJ7X?usp=drive_link - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=199 ### 87. Swarm Auth agent_auth is an OAuth-style authentication and accountability protocol for AI agents. As autonomous agents proliferate — booking flights, drafting emails, moving money — there is no shared infrastructure for answering three urgent questions: who owns this agent, what is it actually allowed to do, and how much should I trust it? The protocol solves this end-to-end. A human verifies themselves and registers their agent by submitting a plain-text skill.md manifest. Claude parses the manifest into structured capabilities, hard constraints, and a recommended trust tier, and generates a human-readable Identity Card. The assigned trust tier is min(ownerVerificationCap, claudeRecommendedTier) — an agent can never inherit more trust than its owner has earned. Any third-party tool integrates via two paths: a lightweight POST /api/agent/signin for backend-to-backend verification, or a full OAuth-style consent flow at /authorize that returns a short-lived JWT bearer token after the owner approves on a hosted consent screen — exactly like "Sign in with Google" but for AI agents. Every action the agent reports through POST /api/events/log is audited live by Claude against its declared scope; out-of-scope or high-risk events trigger an immediate email to the human owner via Resend and stream into a live event console at /dashboard/agent/[id]/events. Agents can be public (browseable trust badge at /agent/[id], for branded products) or private (only verifiable through our certifying API — the Visa-for-credit-cards model, where the merchant can ask "is this real?" without the card needing a public page). The result is a single source of truth that lets humans stay accountable, tools make instant trust decisions, and anomalies surface in seconds. Stack: Next.js 14 App Router, MongoDB, NextAuth (email magic-link), Anthropic Claude (claude-opus-4-5), Resend. - **Team:** [Jonathan Olvera](https://cerebralvalley.ai/u/JohnOlven) - **GitHub:** https://github.com/johnolven/agent-auth - **Demo video:** https://youtu.be/yc5b_JSKvRA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=241 ### 88. Nicolas Arce Every food product hits the same wall: menus are unstructured; PDFs, phone photos, chalkboards, languages, typos, modifiers that look like dishes. Loading them into a system means hours of manual work, restaurant by restaurant. Mise turns any menu into a versioned, searchable dish graph. Claude Opus 4.7 reads the evidence vision-natively (no OCR), extracts canonical dishes with diner-vernacular aliases, attaches modifiers, routes ephemerals and produces an auditable receipt of items dropped by a natural-language filter; one structured-output call per source. A Claude chat answers once. Mise is the system around it: pinned schema, persistent review, JSON ready to plug into a delivery feed, a POS or a voice agent. - **Team:** [Nicolas Arce](https://cerebralvalley.ai/u/NicoA) - **GitHub:** https://github.com/NicoArce10/Mise - **Demo video:** https://youtu.be/ojQpdtRtXe0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=242 ### 89. Balencr AI Helps people understand their personal finance - **Team:** [Richard van 't Land](https://cerebralvalley.ai/u/goldfisher) - **GitHub:** https://github.com/rvantland/balencr-ai - **Demo video:** https://drive.google.com/file/d/1KGDlSQsXbbX2yEbuLKovOzwb-7loVRZi/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=244 ### 90. Kene A dual-surface personal assistant for UK job seekers — **React web app** for deep work and **Telegram bot** for on-the-go. Forward a job URL, get an honest verdict grounded in live UK government data, then ask for a tailored CV, cover letter, salary strategy, or interview prep on demand. Built by someone who spent months job-searching in the UK on a Graduate visa. Every feature exists because a real information asymmetry existed. - **Team:** [Kene Iheanacho](https://cerebralvalley.ai/u/klaustweets) - **GitHub:** https://github.com/DevNick21/trajectory - **Demo video:** https://youtu.be/SRIQH2drsJ0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=247 ### 91. Plan-Audit Plan Audit- Intelligent Plan review and deck builder - **Team:** [Srichandra Kurisetti](https://cerebralvalley.ai/u/Blueprint-IQ) - **GitHub:** https://github.com/permit-iq/Plan-Audit-Hackathon-2026 - **Demo video:** https://youtu.be/gdp1-vgNoqM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=259 ### 92. RESILAND Intelligence RESILAND Intelligence is an AI-powered decision-support platform for the World Bank's $256M RESILAND CA+ landscape-restoration program — covering Kazakhstan, Kyrgyzstan, Tajikistan, Uzbekistan ($153M pilot), Turkmenistan, and Türkiye as a partner geography (drawing on decades of OGM/TEMA dryland afforestation expertise). The problem: Phase-4 feasibility studies for forest-restoration parcels take field consultants 4–6 weeks each. Every day of delay means more desertification, mudflows, and lost ecosystem services. With Opus 4.7 as a teacher model learning from nine real World Bank reference reports, we turned that into 2–3 minutes per report (with prompt caching). Built in 5 days, solo, live in production at resilland.com since Day 2. CORE INTELLIGENCE • Click any of 3,563 cadastral parcels → 9-section feasibility report with passage-level citations to nine reference Phase-4 docs • Drop a field photo → Opus Vision analyzes vegetation health • Drag-drop a KMZ/KML/GeoJSON polygon → spatial scoring + suitability verdict • 5,925 Sentinel-2 scenes, 12-month NDVI per parcel, ERA5 climate, CHIRPS precipitation, 64-species recommendation catalog • 125-document RAG knowledge base (Voyage-3 → pgvector with HNSW index) • Parcel-level Hansen GFC + Köppen-Geiger + WorldClim baselines on all 3,563 parcels • Translate any report EN ↔ TR/RU/UZ, section-by-section, sibling-linked MULTI-TENANT PLATFORM • 5-step setup wizard, Fernet-encrypted credentials in Postgres ("clone, paste your keys, ship" — nothing in .env) • Organizations + role-based invitations (owner / admin / user) • Layer sharing across orgs with per-org pinning • Demo allowance (10k tokens/day, 30-day TTL) lets visitors try without keys ECOSYSTEM • Published MCP server: pip install mcp-resilland — 16 tools that let Claude Desktop / Cursor / any MCP client analyze parcels AND generate full feasibility reports (analyze_area, generate_feasibility_report, translate_report, share-link, PDF) without opening the web UI • WhatsApp bot at +90 535 374 99 71 (chat-only), routing Haiku 4.5 → Sonnet 4.6 → Opus 4.7 in the background • 15 provisioned Managed Agents (7 primary + 8 specialist) — full HMAC-signed bridge + audit trail behind a feature flag • Knowledge Ingestion pipeline: drag-drop DOCX / PDF / PPTX / MD or paste a GitHub repo URL — Opus 4.7 classifies, Voyage-3 embeds ECONOMICS Adaptive cascade routing keeps the bill in single-digit dollars per session: Haiku 4.5 handles ~90% of conversational turns, Sonnet 4.6 drafts report sections with 9× prompt-cache savings, Opus 4.7 escalates only for deep cross-document reasoning and Vision. Stack: Next.js 14 + FastAPI + PostgreSQL/PostGIS + pgvector(HNSW) + Redis + MinIO + Anthropic SDK + FastMCP. Open-source under Apache-2.0 — see ROADMAP.md for the public Q2 / Q3-Q4 / Moonshot kanban (25 issues). - **Team:** [ilhan KILIC](https://cerebralvalley.ai/u/Yapar) - **GitHub:** https://github.com/ilhankilic/resilland-intelligence - **Demo video:** https://youtu.be/S-0B2MfGans?si=gpl6AcmNdetAeouv - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=48 ### 93. OmniBridge OmniBridge is an AI agent that identifies unknown serial device protocols autonomously — a workflow that didn't exist a year ago because it needs Opus 4.7's specific capabilities to work. Plug in a 1980s weighing scale, a Modbus PLC, or an Arduino sensor. OmniBridge runs Claude as a multi-step agent with six purpose-built tools — for binary streams, one surgical call to analyze_binary_structure validates CRC16 across every frame and returns frame structure, enough evidence to commit with high confidence. Every reasoning step streams live. Once identified, Claude-generated regex (ASCII) or byte-offset extractors (binary) run locally at thousands of packets per second, zero ongoing API cost. Traditional protocol integration: engineer reads vendor docs, writes custom parser, ships glue code — eighteen hours median per device. OmniBridge: five minutes. For $1.2 trillion of legacy industrial hardware stranded outside the cloud, that speedup opens entire new markets. This couldn't work with a smaller model. Adaptive thinking lets Claude spend more compute on binary vs ASCII. Multi-turn tool use with preserved reasoning lets the agent pick a surgical tool and commit. Prompt caching delivers ~65% cache hit on repeat investigations, keeping cost near $0.20 per identification — roughly one-thousandth of the engineer time replaced. Open source, MIT licensed, native desktop app on macOS. - **Placement:** Finalist - **Team:** [Adinda Panca](https://cerebralvalley.ai/u/adindamochamad) - **GitHub:** https://github.com/adindamochamad/omnibridge - **Demo video:** https://youtu.be/RX-uGdIQLLU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=51 ### 94. HarryLe (Hau Le) Why it exists: your repo and build are easy for agents to read; the lab often is not—serial consoles, device state, and local tools are out of view. SiliconSmith gives agents and humans that live hardware context in an approval-based way. - **Team:** [Hau Le](https://cerebralvalley.ai/u/Harryle) - **GitHub:** https://github.com/haule98dz/siliconsmith - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=92 ### 95. Bitácora — Universal Memory Index Bitácora is a peer-to-peer knowledge graph where every fact, question, or source you create becomes a cryptographically-signed node — a BIT (Basic Information Timestamp). Memos sync directly between users via WebRTC. No servers store your data. No platform mediates your truth. The graph is reasoned over by Claude Opus 4.7, which detects logical contradictions across distributed memos, maps cross-jurisdictional positions in real-time, and synthesizes consensus from peer-contributed evidence. Managed Agents continue verifying claims autonomously after sessions end, delivering research dossiers anchored to primary sources. The core data structure is engineered for O(1) retrieval (HashTable + Merkle proofs) and offline-first operation. The MVP runs as a React Native Web app — same codebase shipping to iOS, Android, and any modern browser. Cryptographic signatures (Ed25519) and Merkle tree verification mean every memo is tamper-proof and every sync is provable. The problem we solve: today, billions of people fact-check things on their phones, but their findings stay locked on individual devices. No shared truth. No network. No collective intelligence. Bitácora inverts that — your memo network becomes everyone's memo network, but no one owns it. First deployment: Blockchain Summit LatAm Perú (May 13–15, 2026), where 500 institutional attendees, 70 expert speakers, and 10 parallel regulatory tracks generate exactly the dense, contradictory, citation-heavy content Bitácora was built for. Beyond conferences: earnings calls, legislative hearings, academic panels, court proceedings — anywhere distributed truth matters more than platform convenience. We didn't build a better fact-checker. We built infrastructure for collective intelligence without a central authority. - **Team:** [Edward CALDERON](https://cerebralvalley.ai/u/Edwardcadev) - **GitHub:** https://github.com/hashpass-tech/BITACORA - **Demo video:** https://bitacora-1055465030328.us-east1.run.app/ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=131 ### 96. Morris Tung - Intuito Studio BentoDeck is the missing output layer between Claude and Apple's ambient glance surfaces. Claude already reads your systems through MCP, but its answers are trapped inside a chat window — it can't put a number on your iPhone Home Screen, take over your Dynamic Island, or buzz your wrist when something breaks. You say "Show me Stripe MRR, today's signups, and critical errors on my Home Screen." Claude calls BentoDeck's MCP tools and Opus 4.7 runs in nine distinct places across the system: it reads each platform's API docs and emits a verified REST endpoint, generates the JMESPath data transform, picks the widget type, designs WCAG-checked themes from prompts like "make it look like Anthropic", runs anomaly detection on a poll loop with a four-layer cost gate (≈$0.15/user/month vs. ≈$150 naive), explains anomalies in plain English on the Lock Screen, and — for the deepest signals — kicks off a Claude Managed Agents session that writes a multi-paragraph incident runbook the user reads when they tap the banner. Two-tier AI: a fast Messages-API call for the wrist buzz; a long-running Managed Agents session for the deep dive. Native SwiftUI + WidgetKit + Live Activities — no PWA, no Electron. Open-source under MIT. - **Team:** [Morris Tung](https://cerebralvalley.ai/u/intuitostudio) - **GitHub:** https://github.com/intuito-studio/bentodeck - **Demo video:** https://youtu.be/roASE-cP21E - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=134 ### 97. CAIQ The $13 Billion AI orchestration market is currently trapped in "Probability Hell." Incumbent frameworks like LangChain and AutoGen rely on open-ended ReAct loops. At enterprise scale, a single 5% hallucination rate creates a cascading execution decay, making massive multi-agent swarms mathematically unsafe for mission-critical production. SWARM-FORGE is the Defense-Grade Operating System for Agents. I replaced probabilistic guessing with Topological Determinism and Zero-Trust execution. The Architecture: Zero-Shot DAG Compilation: You feed Swarm-Forge a raw enterprise objective. Instead of blindly executing, Opus 4.7 acts as a Meta-Orchestrator, autonomously compiling the problem into an immutable Directed Acyclic Graph (DAG) infrastructure. Mathematical Proofs: Before a single agent spawns, Kahn's Algorithm and a Three-Color DFS validate the execution order. Cycles and infinite loops are physically rejected at plan time. AgentGuard AST Firewall: I engineered a custom 4-stage zero-trust middleware. If an agent hallucinates a rogue command, my ActionFirewallVisitor drops the capability at the Abstract Syntax Tree (AST) level—before the code ever reaches the interpreter. Byzantine State Lock: A node only passes if it achieves a Byzantine consensus threshold verified by a fail-closed Sonnet 4.5 Reward Judge, ensuring 0.0% unauthorized execution in my benchmark suite. Swarm-Forge doesn't just route agents; it provides the deterministic, mathematically secure rails required to safely scale the $13 Billion orchestration market. - **Team:** [MOIZ SIDDIQ](https://cerebralvalley.ai/u/moiz_siddiq) - **GitHub:** https://github.com/f2025408135-cyber/SWARM-FORGE - **Demo video:** https://www.youtube.com/watch?v=N8iFqA344jo - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=144 ### 98. Vladislav Popov Claude Habitat is a Windows-desktop application (.exe) that turns Claude into a self-architecting personal AI workspace. No terminal, no config files, no technical knowledge required. The app runs a 5-step wizard: enter your API key, answer a few questions about yourself, and Claude builds a personalized workspace inside your messenger. The system is messenger-agnostic by design (Discord, Slack, WhatsApp, Telegram) - for this hackathon we ship with Telegram. The workspace includes topic-based channels, file processing, persistent memory, and semantic search. What makes it unique: • Self-architecting: Claude analyzes who you are (role, tools, workflows, pain points) and generates a custom set of topics and capabilities. A bookkeeper gets VAT, Payroll, Client Books topics. A developer gets different ones entirely. • Zero-config deployment: One .exe, no installation, no dependencies. Works on any Windows machine. • Real file processing: Send PDFs, images, documents to any topic - Claude reads, summarizes, and saves them to local folders on your disk. • Persistent vector memory: Every message is indexed in LanceDB. Use /search to find anything across all topics. • Fully local: All data stays on the user's machine. The bot runs as a child process inside Electron. Built entirely with Claude Code (Opus 4.7) over 5 days. The Electron app, Telegram bot, onboarding system, habitat builder, and vector memory were all written by Claude. The pixel-art mascot (Claude Toad) was animated in code by Claude - eyes track the cursor, random blinks, speech bubbles with personality. Click him and he croaks; keep clicking and he starts speaking in the developer's actual voice. There's a samll hidden easter egg if you're persistent enough. UI mockups for the demo video were created in claude.ai/design - **Team:** [Vladislav Popov](https://cerebralvalley.ai/u/devladpopov) - **GitHub:** https://github.com/devladpopov/claude-habitat - **Demo video:** https://youtu.be/vP-ILPegrvA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=192 ### 99. Danyar Group GENESIS transforms academic research papers into interactive, runnable prototypes in under 60 seconds. The problem: researchers publish thousands of papers weekly, but almost none include working code. Understanding a paper means manually decoding equations and rewriting algorithms from scratch — a process that takes days or weeks. GENESIS solves this with a single URL input. Users paste an arXiv link, and Claude Opus 4.7 extracts the paper's "Paper DNA" — title, classification, core equations, tunable parameters, and optimal visualization type. Then a second-stage high-effort prompt generates the actual algorithm as pure JavaScript. The code runs instantly in a sandboxed Sandpack executor with interactive parameter sliders that update the visualization in real-time. We also built an AuditWorker using Anthropic's Managed Agents that automatically verifies the paper's numerical claims against the generated code, plus a Debate mode where two Opus 4.7 instances evaluate code fidelity. - **Team:** [Danyar Othman Azeez](https://cerebralvalley.ai/u/Danyar82) - **GitHub:** https://github.com/danyar82/Genesis_Hackathon - **Demo video:** https://youtu.be/EzUpqeKsKHQ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=193 ### 100. Florian An app that will allow anyone, no matter their background, to master foundation of Python. - **Team:** [Florian Muljono](https://cerebralvalley.ai/u/FM125) - **GitHub:** https://github.com/FlorianMuljono/pyquest - **Demo video:** https://drive.google.com/file/d/1Ja1cdn-nKX2Me-40brdDBB87UFSqLfaq/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=195 ### 101. rhmoller A kid raises a hand to the webcam and a puppet comes to life — mouth, eyes, and gestures driven by their fingers via MediaPipe. An AI puppet on stage listens with words AND body language, and replies in character. When the kid says "let's go to the beach", a sun, a sand castle, and a beach ball appear. When they say "I want sunglasses on my puppet", sunglasses appear. When they say "I want a watermelon hat" — something the catalog doesn't have — the AI says "ooh, let me dream that up!" and a few seconds later a watermelon-shaped hat fades onto its head. Asked again, it pops in instantly from cache. Behind the scenes, two Claudes work in parallel. A conversation Claude performs the live character with emotion, gaze, and gestures. A separate Claude designs each novel prop on the fly from a tight set of THREE.js primitives. Its compositional reasoning is what makes the co-creation feel real-time. An interface for play that didn't exist a year ago. - **Placement:** Finalist - **Team:** [Rene Hangstrup Moeller](https://cerebralvalley.ai/u/rhmoller) - **GitHub:** https://github.com/rhmoller/virtual-puppet-theater - **Demo video:** https://youtu.be/qLuGU4PQNss - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=201 ### 102. Tharshi I run a produce wholesaler. Every morning my team has to buy around 50 - 200 different items from 25+ suppliers at the Ontario Food Terminal. The price lists come in as PDFs and scans. The relationships with each supplier live in iMessage. A lot of the job is just remembering things. BuyOps takes that morning and gives it to a team of Opus 4.7 agents. One agent reads every price list and turns the mess into one clean index. Then for each item we need to buy, a separate Claude session looks at the prices, reads the last two weeks of texts with that supplier, and picks the best place to order. If something is off, like a supplier shorted us yesterday or a price quietly went up, the agent stops and flags it instead of guessing. When the picks are in, it groups them by supplier and writes the actual message you would send. You hit copy, paste, edit (if needed) and send. I built it on real data from my own company, with all the names changed. What used to take a full time data engineer and hours of operational work per day, the agents do in about 45 minutes. - **Team:** [Tharshi Sri](https://cerebralvalley.ai/u/tharshi) - **GitHub:** https://github.com/tharshi92/buy-ops - **Demo video:** https://www.youtube.com/watch?v=HVcN36nUrQs - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=238 ### 103. Anton Abyzov Skill Studio is a local-first IDE for AI skills. Author once, install everywhere — Claude Code, Cursor, Codex, Windsurf — with npx vskill install . It ships a security-scanned marketplace of 100K+ skills, an in-app generator that runs on any model (Opus 4.7, GPT-5, Gemini, local LM Studio via BYO), auto-generated test cases, GitHub-CLI-powered publish flow that creates the repo for you, and SSE-pushed live updates the moment a skill changes. Solves the lock-in, drift, and "is this still working?" gaps that Anthropic's SKILL.md spec leaves open. - **Team:** [Anton Abyzov](https://cerebralvalley.ai/u/EasyChamp) - **GitHub:** https://github.com/anton-abyzov/vskill - **Demo video:** https://youtu.be/TM_UwAXVLg4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=243 ### 104. Friday Deployments astify is a CRM storage and backup platform built for Salesforce. Salesforce charges around $250 per gigabyte per year for native file storage — a medium org with a few terabytes of attachments burns six figures every year just to keep that data sitting there. Vastify sits invisibly between Salesforce and a cloud bucket the customer already owns (S3, GCS, Azure, R2). It proxies the same OData endpoint that every dependent Salesforce app reads from, so apps don't notice the data moved — but the actual files now live in customer-owned, AES-256-encrypted, tier-routed object storage at roughly 1/100th the cost. Files stay queryable, attachments stay accessible, the bill drops by an order of magnitude. Three AI agents power the day-to-day: Setup Agent — installs Vastify in 42 seconds via 6 typed tool calls. Inspects the org, picks the cheapest backend for the data shape, deploys the managed package, verifies the OData endpoint. Diff Explainer — reads the change-set when you need to restore from a backup and tells you exactly what's safe to bring back, what would overwrite live work, and what would corrupt the schema. Rule Generator — turns plain-English routing intent ("move all contracts older than 90 days to Glacier and ping the AE in Slack") into validated JSON routing rules. All three are built on the Anthropic Agent SDK with Claude Opus 4.7. Live demo in your browser. - **Team:** [James Collard](https://cerebralvalley.ai/u/Exotic52) - **GitHub:** https://github.com/Exotic209093/Vastify - **Demo video:** https://youtu.be/_OcVUofbFTM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=49 ### 105. Beyond The Limit (BTL) My father has Alzheimer's. This submission is the tool I wished existed for him. Kith holds the person, when memory cannot. 55 million people are losing themselves to dementia right now. By 2030, 110 million+. Every product built for them solves logistics — pill reminders, GPS, alerts. None of them hold the person. Kith does. It's a 24/7 AI companion for patients, families, and care teams — designed around four Opus 4.7 capabilities that earlier LLMs could not have supported. A SHOEBOX BECOMES A LIFE STORY. Most families freeze at a blank onboarding form. They have photographs. Kith uses Opus 4.7's 2576px vision to read scanned family photos — including the handwritten captions on the back, signage in the background, and text in cards and letters — drafting a structured Personhood Map the family reviews. Names, dates, relationships, occasions. Every claim flagged with its source image. multiple photographs in, few seconds later, the family has a simple document of their loved one's life. A NURSE'S NOTEBOOK ACROSS EVERY SESSION. Kith keeps five living markdown files in file-system memory — today.md, recurring_themes.md, joy_log.md, concerns.md, gentle_boundaries.md. Read at the start of every session. Updated after every turn. Restructured nightly by a reflection agent. The patient forgets yesterday. Kith does not. This is what Opus 4.7's improved file-memory was built for. A SAFETY ARCHITECTURE, NOT A SAFETY DISCLAIMER. Every patient response runs a silent self-verification pass: list every specific claim, identify its source, rewrite if unsourced. Dissonant-data resistance handles the hardest moments — when the patient confidently asserts a fabrication, Kith holds ground without harshly contradicting. Adaptive thinking responds in 1 second on "who is this?" and pauses for 4 seconds on "where is George?" when George has passed. In this domain, a hallucinated memory is not a product bug — it is a harm to a person who cannot correct it. A WEEKLY LETTER, NOT A DASHBOARD. Five parallel Claude subagents — Mood, Memory, Changes, Routines, Joy — read each week's conversation log through their specialized lenses. A sixth synthesizes them into ~400 words written like a thoughtful nurse, not a notification. The kind of letter a family keeps. Every Opus 4.7 capability we used is load-bearing. Every safety rule is architectural, not aspirational. This is what AI companion care actually looks like when you build it for someone you love. - **Team:** [M Noman](https://cerebralvalley.ai/u/nomang) - **GitHub:** https://github.com/nomang/kith--for-claude-4.7-hackathon.git - **Demo video:** https://www.loom.com/share/7d25029355da4a26b624b6f8c072b84d - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=94 ### 106. Mohsen Azimi tsz is a fast and eventually more sound TypeScript type checker. I read in the rules that existing projects do not qualify but I explicitly mentioned that I will be working on this for the hackathon when I signed up. I'm okay if this project is not accepted. It is still a work in progress so it was neither started or completed during this hackathon. But I would love to get more publicity and hopefully contributions to the project anyway. I've been working on this since Jan 1st! - **Team:** [Mohsen Azimi](https://cerebralvalley.ai/u/mohsen1) - **GitHub:** https://github.com/mohsen1/tsz - **Demo video:** https://tsz.dev - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=133 ### 107. jbeans Quorum is a control plane for safe, auditable, policy-gated AI agent execution on code and infrastructure. Instead of letting a single overprivileged agent mutate systems directly, Quorum turns actions into structured findings and typed proposals, runs policy checks, requires quorum and human approval for high-risk changes, executes through real GitHub and Fly.io actuators, verifies health afterwards, and rolls back automatically on failure. The system keeps a tamper-evident append-only event log and exposes a live operator console so teams can inspect exactly what happened and why. It solves the trust gap in agentic engineering: current coding and infra agents can take actions, but they rarely provide strong governance, auditability, or safe recovery when something goes wrong. - **Team:** [Jayden Piao](https://cerebralvalley.ai/u/jaydenpiao) - **GitHub:** https://github.com/jaydenpiao/quorum - **Demo video:** https://drive.google.com/file/d/1toXH542fBH1VClpWig8YPXIfvP7G3HY1/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=136 ### 108. ECBC Huantuk is : 1. Diagnosis synthesizer : input health reports and clinical history, after first understanding, answer further clinical history and input further health reports (as recommended and other new health reports), output diagnosis and current clinical practice guideline (CPG) management 2. Drug discovery : input huantuk if unresolved after CPG management, read through all personalized health datas, research through medical publications, clinical trials and related evidence based articles, output repurposed drug, off-label indicated medication and applicable clinical trials - **Team:** [Tagenda Cemerlang](https://cerebralvalley.ai/u/ECBC) - **GitHub:** https://github.com/RC24bc/huantuk - **Demo video:** https://youtu.be/hJgtMu-3_rs - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=140 ### 109. Team MinistryXR Engineering Ops is an AI-powered QA and Security governance platform built on Claude Managed Agents. As an AI-powered XR studio, our developers at MinistryXR ship faster than ever using AI coding tools. But our QA and security processes couldn't keep up. Outsourced reviews were costly, manual security audits took days, and compliance mapping was done by hand. Engineering Ops uses five specialized Claude Managed Agents working in parallel to automate QA testing, security scanning, dependency analysis, and code fixing all with human approval at every critical step. Every finding is automatically mapped to PDPA, CCoP 2.0, and ISO 27001 compliance clauses. Supervisors assess findings, sign off, and maintain a full audit trail. The Auto-Fix agent writes tests first and only pushes to GitHub when all tests pass. Tested live on our actual MinistryXR codebase — 33 issues found in under less than 25 minutes. What previously took a week of outsourced review now happens in minutes. - **Team:** [Avin Mahajan](https://cerebralvalley.ai/u/avin17) - **GitHub:** https://github.com/skyofreality/engineering-ops - **Demo video:** https://youtu.be/hKIxDry5hoI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=141 ### 110. Ali Mahmood Local independent newsrooms have a sales problem most readers never see: they can't afford sales people. Sales for news media is genuinely expensive expertise (pricing, editorial-firewall navigation, prospect research, partnership reviews), and most of an experienced salesperson's time gets eaten by grunt work: assembling media kits, building sell sheets, drafting prospect research briefs for every meeting. The pragmatic answer the industry is converging on is the shared-services hub: one experienced salesperson covers four to ten outlets at once. That only works if the grunt work is automated. Otherwise that person burns out across ten clients in six months, and the small newsrooms lose the only sales function they could afford. Sponsor Desk is that automation. It connects to an outlet's first-party data and drafts the three documents (media kit, per-product sell sheets, and a prospect research brief) in seconds using Claude Opus 4.7. The structured output schemas carry a top-level RefusalEnvelope branch, so the editorial firewall is enforced at the schema level. A demo brief on a fictional Portland utility that filed a rate increase the demo outlet is independently reporting on flags rate filings and grid reliability as editorial-conflict topics, with named specifics pulled from the utility's own source files. Every generated document writes a provenance sidecar with cache telemetry and pinned content hashes for the editorial-safety block, the prompt contract, and the document-design Skill. The audit trail is the answer to "where did this number come from?" Today's data is synthetic. The architectural hook for real data is already in the codebase: data_sources/ga4_stub.py exposes every method's signature ready for v2 adapters into GA4, Mailchimp, Stripe, and Piano. v2 is where this genuinely helps a real newsroom: real outlet data flowing through the same firewall, one experienced salesperson sustainably covering several newsrooms, professional sales for outlets that otherwise couldn't afford it. MIT-licensed. Self-hostable. Six phases shipped end-to-end. 68 tests passing. Real-API verified. - **Team:** [Ali Mahmood](https://cerebralvalley.ai/u/am83456) - **GitHub:** https://github.com/hockeystick/sponsordeskplus - **Demo video:** https://youtu.be/vAY0_pLmud8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=143 ### 111. Month Proof Month Proof automates month-end close for US finance teams. Upload your financial Excel files → the agent reads them, compares to history, flags anomalies, writes a plain-language report, and sends it via email — all in under 2 minutes. Finance teams spend 10-15 hours/month on manual variance analysis and narrative writing. Month Proof reduces that to zero human-hours for the first draft. The system uses a 3-agent pipeline: Parser (column mapping with Haiku), Comparison (pure pandas variance calculation), and Interpreter (narrative generation with Opus 4.7). The critical innovation is the numeric guardrail: Claude never does math — all calculations happen in Python, and a verification layer ensures the narrative matches the pandas output within 2% tolerance. No report is saved until the guardrail passes. Built with FastAPI, React, Supabase, and Claude Opus 4.7. Tested with real GL exports from NetSuite, QuickBooks, and manual consolidation scenarios. The entire codebase was developed using Claude Code as the primary development engine. - **Team:** [Can Soganci](https://cerebralvalley.ai/u/Janso) - **GitHub:** https://github.com/jansoganci/iron-ledger.git - **Demo video:** https://www.loom.com/share/93db0269afaa400aa612fe0311291094 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=149 ### 112. MaestrIA My dad is a master carpenter in Chiloé, southern Chile. 30 years on the job, most of them as site foreman, and 8 of those years restoring heritage churches with the Fundación Iglesias Patrimoniales de Chiloé. I am his apprentice and I also run the development and business side of his company, Constructora Patrimonial Chiloé. I built its website (constructorapatrimonialchiloe.cl) with Claude Sonnet a few months ago. The website helps people find him. It does nothing to help a client know whether his quote is fair, whether his diagnosis is correct, or whether the trade showing up is the right one for the problem. In Chile alone, 280,000 construction workers are in the same position (INE, 2025). MaestrIA is what I built so a regular family can get an architect-grade diagnosis for the kind of repair an architect would never come to see. You take a photo, describe the problem in your voice in Chilean Spanish, and say where you live. Claude Opus 4.7 thinks live on screen, draws boxes over the damage as it reasons, names the cause, estimates the work in real units (half a day, two days, not "3.5 hours"), and pulls together a committee of trades that argue on screen until they agree on what to do. A managed agent checks the prices against Chilean retail. A second agent writes the WhatsApp message to the chosen maestro for you: greeting by name, problem summary, materials, ballpark budget, time estimate, link to the full diagnosis. The client stops arriving with "my wall got wet" and starts arriving with a brief. The whole system was measured against my dad. He diagnosed 12 real photos in audio, without seeing the AI. Across 9 dimensions (pathology, material, days of work, CLP cost, severity, honesty under uncertainty, and others), MaestrIA agrees with him 81% of the time. Before integrating his interview into the knowledge base, the score was 74%. Same model, same prompt, same photos. The +7 points are his. - **Placement:** Finalist - **Team:** [Benjamin Torralbo](https://cerebralvalley.ai/u/Benja) - **GitHub:** https://github.com/Benjxyg/MaestrIA - **Demo video:** https://youtu.be/rkH4AjoTL5Q - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=194 ### 113. Lian Passmore Bea is a voice-first coach for whānau and families, built during the Built with Opus 4.7 hackathon. Most family apps ask people to log moods, track habits, or fill in forms. Families don't have time. Bea listens instead. She joins everyday family kōrero (conversations), transcribes, and reflects in the moment with warmth. The thinking happens between sessions. Three Opus 4.7 reasoning agents do work a smaller model would visibly fail at: a group session reader handles multi-speaker transcripts and per-member summaries; a pattern detector reasons across weeks; and a Coach decides what Bea should gently raise next, or leave alone. Five of eleven guardians use Opus 4.7, selectively, where extended thinking earns its cost. Bea's coaching stance comes from Takitaki mai (the Māori adaptation of Motivational Interviewing) and mana-enhancing practice. She borrows the skills. She leaves the rituals to the people who own them. Built by Lian Passmore, Aotearoa New Zealand. - **Team:** [Lian Passmore](https://cerebralvalley.ai/u/pacificviking) - **GitHub:** https://github.com/lianpassmore/bea - **Demo video:** https://youtu.be/qUb3TwUUQLg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=198 ### 114. Filipe Burmester Modern medicine is prescribing more drugs to the same patient than ever before. ~42% of U.S. adults aged 65+ take five or more prescriptions, and adverse drug reactions are a top-six cause of preventable death — the JAMA estimate is ~106,000 fatal ADRs per year in U.S. hospitals alone. Two failure modes drive most of that harm: the cytochrome P450 enzymes (which metabolize ~75% of clinical drugs and vary genetically between patients), and drug–drug interactions that compound silently across a regimen. Pharmacyp is a clinical decision support tool that catches both. A clinician enters a patient's drug list and CYP genotype; Pharmacyp returns a structured risk report through three layers: a deterministic rule engine over CPIC pharmacogenomic guidelines and curated DDI data (DDInter 2.0), two graph-neural-network microservices that predict interactions and CYP behavior from chemical structure when no curated data exists, and a Claude Opus 4.7 explainability layer that annotates every finding with a citation back to the underlying clinical literature. Every output is labeled by source — RULE-BASED, ML-PREDICTED, or INSUFFICIENT DATA — so uncertainty is never hidden behind a confident tone. - **Team:** [Filipe Burmester](https://cerebralvalley.ai/u/filipeburmester) - **GitHub:** https://github.com/fburmester/pharmacyp - **Demo video:** https://youtu.be/-fyEVTyoO9s - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=202 ### 115. Felipe M Affonso Lens is an open-source consumer welfare agent for the AI shopping era. Paste any AI shopping answer (from ChatGPT, Claude, Gemini, Perplexity, or Rufus), drop in a retailer URL, or shop with the Chrome extension running, and Lens audits the recommendation in real time. It flags every confabulated spec the AI cited (the "stainless-steel housing" that's actually plastic, the "15-bar pressure" that's a marketing number), surfaces the spec-optimal product the AI skipped, names dark patterns at checkout (preselected fees, drip pricing, fake scarcity, forced continuity) with the FTC junk-fee rule cited inline, and drafts the refund or complaint letter for your one-click approval. After you buy, the watchers keep working. CPSC and NHTSA recalls, retailer price-drop windows, subscription pre-charges, and firmware CVEs cross-match against your purchases, and Lens drafts the intervention before you ask. One Cloudflare Worker powers a web app, a Chrome MV3 extension, a mobile PWA, an MCP server with 16 tools callable from Claude Desktop, and a public OpenAPI surface. The deterministic ranker stays language-model-free once your utility function is formed, so the top pick is reproducible from the utility function. MIT licensed, no affiliate links, $0 affiliate revenue. Built with Opus 4.7. Live at https://lens-b1h.pages.dev. - **Team:** [Felipe Affonso](https://cerebralvalley.ai/u/felipemaffonso) - **GitHub:** https://github.com/FelipeMAffonso/lens - **Demo video:** https://www.youtube.com/watch?v=Rh4eufbcD3I - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=245 ### 116. gg Match the chaotic brilliance of a human to the strategic, operational nature of AI. Davaut is a context-management substrate that ends two things at once: AI hallucination and the 100k-token context dump. Today's AI tools have the memory of a goldfish — you paste your life in, the model wades through all of it to answer one focused question, you start fresh tomorrow, repeat. It's expensive, shallow, and forces you to be the memory system. Davaut turns your work — docs, chats, code, plans, email — into a graph of nodes organized by layers (relations, tags, heat, importance, session timelines) treated as composable dimensions. A workspace is a canvas. The Librarian, an Opus 4.7 Managed Agent, composes those layers like dimensions, assembles a course (a tight, mutable bundle of just the nodes that canvas needs), and serves it. Overnight, Dream — a second Managed Agent — consolidates: decays heat, promotes important tags into nodes, generates summaries. Your second brain gets smarter while you sleep. The result is threefold. Cheaper — you stop burning tokens on irrelevance. Smarter — the model goes deeper because it isn't drowning. Grounded — the Librarian only returns nodes that exist, so the AI on top physically cannot hallucinate. Every claim anchors to a real node with a real provenance trail. Davaut plugs into the SaaS you already use — Google Docs, Notion, email, your repo — no migration, originals sealed forever in the immutable Raw chunk. It's built for anyone running parallel threads: the hacker juggling architecture and three half-formed ideas across linked canvases that inherit each other's context, and the founder who needs the ultimate secretary that pulls exactly the right context the moment it's needed. Davaut is the memory under the model. - **Team:** [Ray Tan](https://cerebralvalley.ai/u/Raytan) - **GitHub:** https://github.com/101011101/Davaut/tree/ultra-mvp/davaut - **Demo video:** https://www.youtube.com/watch?v=m1O1Yj3NHbY - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=248 ### 117. Skepya very year, 154 million people suffer a stroke. A third develop aphasia — they recognize the world around them but can't find the words to describe it. A mechanic who spent forty years naming every tool in his garage suddenly can't say "wrench." Therapy helps, but it happens twice a week. Between sessions, patients get generic flashcard apps that treat everyone the same. The best tools cost $250/month with zero personalization. RehabVox builds personalized therapy games from the patient's own life. A therapist uploads photos from the patient's real environment — their garage, kitchen, favorite motorcycle — alongside clinical data and personal details: hobbies, passions, even inside jokes. From there, a single AI handles everything. It identifies and extracts objects from the photos, designs a complete therapy game with exercises, progressive hints, and personalized feedback — then generates every audio instruction in a cloned familiar voice from a short family recording. The patient opens one link and plays. Zero AI at runtime — everything is pre-generated. After each session, the AI analyzes performance, identifies weaknesses, and rebuilds the next session to target them. The therapist reviews and approves at every step. This approach is grounded in research: aphasia patients recover significantly faster practicing with familiar objects. Professional vocabulary is the most resistant to brain damage — a mechanic who can't say "pen" might still say "wrench." RehabVox starts there. The entire pipeline runs locally with open-source models. No cloud GPU required. Objects are extracted once and reused — regenerating a new game takes seconds. Cost per session: fifteen cents. Between therapy sessions, there's finally something. - **Team:** [Rares Roscan](https://cerebralvalley.ai/u/Rarrs) - **GitHub:** https://github.com/roscanrares/Claude-Hackathon - **Demo video:** https://drive.google.com/drive/folders/1XYYO7R45eopuXZi4hv1KSciy4vGUvRUo?usp=share_link - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=264 ### 118. AIDEN PACT is a semantic compression format that beats ZIP by 40% on codebases and replaces Claude Code's /compact with an 18.4x lossless compaction engine. Three things in one tool: 1. File compression — brotli solid archive that concatenates all files into one compression stream. Cross-file pattern deduplication that ZIP physically cannot do. 40% smaller on a 79-file real project. Semantic inspection shows every function, import, and class without decompressing. 2. Claude Code compaction plugin — one command (pact install --global) hooks into every session. Heuristic structural extraction at 50% context, zero API calls. 18.4x compression vs /compact's 4.0x, with 92% information retention vs 68%. The AI keeps structured state instead of lossy prose summaries. 3. Native OS integration — right-click any file on macOS, Windows, or Linux to pack/unpack. Native Swift progress UI on macOS. One command setup: npx pact-cc setup. BARUTU SNAKE: PACT compressed the Claude Code session that built PACT. 179 turns, 1.6M tokens → 46K tokens, 34.3x lossless. Reproducible: node benchmarks/run.mjs --tasks 012-barutu-snake. 150 tests passing. Built in 5 days. pact.zip - **Team:** [Aiden Hecker](https://cerebralvalley.ai/u/aidenhecker) - **GitHub:** https://github.com/hmatrades/PACT - **Demo video:** https://youtu.be/No8pofFf8-U - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=266 ### 119. Ryan MacDonald Millions of servers run the world's critical infrastructure, defended by teams too small to keep up. They've stitched together 15 years of open-source tooling but attackers now weaponize CVEs in hours while defenders are still grepping logs. Blacklight is a low-dependency agentic harness that turns the hooks already wired into existing OSS security tools into an autonomous defense layer. Detect on one host, triage and contain across the fleet, at attacker speed. The substrate isn't migrating to enterprise XDR. Blacklight makes the substrate agentic. - **Team:** [Ryan MacDonald](https://cerebralvalley.ai/u/rfxnryan) - **GitHub:** https://github.com/rfxn/blacklight/ - **Demo video:** https://youtu.be/cksgujyY7Ms - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=269 ### 120. Brownfield Code Surgeon # Brownfield Code Surgeon ## The Problem Statement AI coding assistants are great on greenfield projects but not-so-great on brownfield ones — the sprawling, undertested, half-documented codebases. Unfortunately, the brownfield code makes up most real software work. AI assistants tend to quietly break things nobody had tests for, in such old codebases. ## The Approach Brownfield Code Surgeon fixes that with a **seven-phase workflow** — Plan → Map → Break → Cover → Implement → Refactor → Finish. The roots go back to Michael Feathers' *[Working Effectively with Legacy Code](https://objectmentor.com/resources/articles/WorkingEffectivelyWithLegacyCode.pdf)* (2002) — seams, characterization tests, the whole vocabulary humans have used on legacy code for two decades. I, along with my colleagues, translated that workflow for agentic use in a [paper](https://doi.org/10.5281/zenodo.19640171) published April 2026. This repository submission is the paper's productionized form. Each phase has its own subagent, its own approval gate, and a set of forbidden moves the agent can't bypass. ## What's in the Submission? During the hackathon, I have created a Claude-code-focused package with one backbone, three ways to use it, all writing the same artifacts: - **Operating Theater UI** — Picture a live operating theater. A cute cat that represents your codebase is on the operating table, breathing, twitching, reacting in real time as Claude operates on your legacy codebase. Vitals tick across the monitor — coverage, seams, failing tests. Every incision the agent makes shows up on screen. That's the UI we have in our submission. - **Claude Code plugin** — slash command `/surgery`, one subagent per phase, hooks that block dangerous edits - **SDK runner** — a Node CLI that drives the same pipeline through the Claude Agent SDK - **Managed agents hand-off** — The Finish phase can hand off to **Claude Managed Agents** so long-running cleanup doesn't tie up your laptop. ## Why Opus 4.7? Opus 4.7's 1M context is doing real work here: a single subagent can hold an entire legacy module in view and find seams a human reviewer would skim past. The seven-phase split keeps every decision auditable. ## Proof It Works PR #33 vs PR #34 on `vivganes/kanbanstr` github repo — same brownfield task, with-surgeon vs without — tighter diffs, real coverage, no forbidden moves. Coverage improved from 0.82% to 19.24%. ## Licensing & CI Open source under MIT, 90% line coverage enforced in CI. - **Team:** [Vivek Ganesan](https://cerebralvalley.ai/u/vivganes) - **GitHub:** https://github.com/vivganes/brownfield-code-surgeon - **Demo video:** https://www.youtube.com/watch?v=ypbRUI6m5jc - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=50 ### 121. Vivax Clinical research leads spend weeks manually identifying care gaps and overprescription patterns in their patient population — querying EHRs, cross-referencing guidelines, validating statistically, packaging interventions. The Clinical Hypothesis Scout collapses this loop. A 7-phase Claude Opus 4.7 agent runs over a Neo4j graph of 21,582 patients, four medical ontologies (SNOMED, RxNorm, ATC, LOINC), and 458 recommendations extracted from 64 American clinical guidelines. Each surfaced hypothesis must pass three statistical guardrails, survive a self-critique that auto-declines synthetic-data artifacts, and arrive packaged with five live PubMed citations and an operational action plan covering intervention, owner, complexity, and dollar impact. What separates it from a hypothesis generator is the continuous-monitoring layer. The scout fingerprints each clinical pattern and emits NEW alerts when a pattern first appears, STRENGTHENED when its cohort grows by 20% or more across batches — in the demo dataset, a bisphosphonate-overprescription signal grew from 303 to 450 to 587 patients across three consecutive batches, the kind of consolidating pattern a research lead would want paged about. A Next.js dashboard exposes kept and declined queues with reasons, a cohort browser, a streaming NL query interface, and reproducible Cypher for every finding. - **Team:** [Umut Kilinckaya](https://cerebralvalley.ai/u/Umut-K) - **GitHub:** https://github.com/Umut-K/GraphDB_for_clinical_records - **Demo video:** https://youtu.be/cuB3ykZn_nE - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=52 ### 122. TrayStones ARIA captures the floor operator's knowledge and puts it to work. Drop in the manufacturer's PDF, answer some calibration questions, and you're live in ten minutes. After that, ARIA continuously ingests everything that already exists — live signal trends, operator logbook entries, shift notes, machine failure history, and computed KPIs (OEE, MTBF, MTTR) — building a knowledge base that grows with every incident. When something goes wrong, five agents pass the problem like a real maintenance team passes a ticket: detection → diagnosis → work order → memory. Five Agents : - KB Builder — reads the manufacturer PDF with Opus 4.7 vision and captures the operator's floor knowledge through a short calibration dialogue. Live before they leave the terminal. - Sentinel — watches live signals against the KB; detects breaches, forecasts signal tails, and judges whether a drift warning is worth surfacing. - Investigator — the centrepiece. Diagnoses anomalies with Opus 4.7 extended thinking; writes and executes Python in Anthropic's sandboxed container to compute exact degradation rates from raw signal data; recalls and builds on every past failure. - Work Order Generator — turns the RCA into a printed sheet: root cause, remediation steps, exact part number, intervention window. - Q&A — natural-language operator chat; hands off to the Investigator when a deep diagnosis is needed. All five agents share 17 MCP tools as their only path to the database. Generative-UI artifacts — charts, diagnostic cards, work orders — stream into the operator's chat as the agents work. - **Placement:** Finalist - **Team:** [Idriss Benguezzou](https://cerebralvalley.ai/u/zestones), [Adam HNAIEN](https://cerebralvalley.ai/u/vgk) - **GitHub:** https://github.com/zestones/Aria - **Demo video:** https://youtu.be/Hen24w2Jyz4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=95 ### 123. Praktikal The problem. Behind every great science lesson is a teacher who built it. In Estonia, that means roughly fifty hours a week — fifteen above the legal limit. Ninety-two percent burn out; a third of new teachers quit within three years. The hardest hit subjects are physics, chemistry and biology. Existing platforms let teachers share slides and worksheets, but a physics lesson isn't a slide — and a play-only video isn't a lab. There's a gap between "explain it" and "let the student do it." What we built. praktikal-sim is a simulation creator built for the classroom. A teacher describes what they want to teach in plain English; an agent assembles a real, physics-accurate simulation on planck.js that runs in any browser. The agent reasons out loud — picking bodies, setting gravity, wiring springs, tuning parameters — and the resulting sim is something a student can drag, throw, restart, break. Sliders turn variables into questions: what happens if the mass doubles? what if friction is gone? How it slots in. Sims drop straight into a Praktikal lesson, or share via a link. Same scenario, every student, but each one runs their own experiment — exactly the inquiry cycle (observe → hypothesize → experiment → analyze → conclude) science teaching is supposed to be, without the equipment, breakage and cleanup. Stack. TypeScript pnpm monorepo. Vite + Vue 3 frontend. planck.js physics with an engine-agnostic sim.* seam. Agent authoring lives behind a small API. Deterministic, seeded runtime so every share-link replay is reproducible. Built from teachers, for teachers — by Praktikal, the team behind Estonia's hands-on STEM platform. - **Team:** [Omari Loid](https://cerebralvalley.ai/u/Aclimate) - **GitHub:** https://github.com/Praktikal-Education/praktikal-sim - **Demo video:** https://drive.google.com/file/d/1GytTqw9mpKuYQUA06BAb7BSkIzdvAgCV/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=139 ### 124. Ryze Pit Stop is a mobile-first AI vehicle diagnostic assistant that helps drivers understand car problems before visiting a mechanic. Instead of giving generic advice, Pit Stop runs a guided investigation. Users describe a symptom, such as brake grinding or an engine noise, and the app asks targeted follow-up questions to narrow down the issue. It then uses an interactive car model with hotspots to show exactly where the user should inspect or take photos. The app validates uploaded evidence, rejects unclear images, and guides the user toward better photos when needed. Once enough information is collected, Pit Stop uses AI to generate a clear diagnostic report with likely causes, urgency level, safety guidance, and recommended next steps. Each vehicle has its own saved history, so users can track past issues, revisit diagnoses, and ask follow-up questions. The goal is to make car troubleshooting clearer, safer, and less intimidating for everyday drivers. - **Team:** [Fateen Ajaz](https://cerebralvalley.ai/u/fateen) - **GitHub:** https://github.com/fateenajaz/pitstop - **Demo video:** https://youtu.be/XzGkm49Um7c - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=197 ### 125. Team Dax Causalist makes any codebase legible — to humans and to agents. Drop in a GitHub URL and four Opus 4.7 agents (Structure, Dependency, Semantic, Oracle) draw your repo as a live causal graph in 30–60 seconds. Files become nodes, imports and calls become edges, every node is classified by what it actually does (UI, API, logic, data, infra). For a beginner staring at an unfamiliar repo, it replaces hours of clicking through folders with a single picture of how the thing fits together. For a senior dev, it's the mental model you'd build over a week, ready in under a minute. But the bigger unlock is for coding agents. Today every agent operates blind: open one file, edit it, run tests, watch them fail, grep, repeat — burning context on rediscovery. Causalist gives agents a causal substrate. Through11 CLI tools (blast_radius, find_path, get_neighbors, summarize_node…), an agent can ask "what does this change break?" before changing anything, follow real cause-and-effect chains across files, and load only the slice of the repo that actually matters — slashing token usage and stopping the trial-and-error loops that eat sessions alive. Tree-sitter AST verification marks every LLM-inferred edge as confirmed or unverified, so the graph the agent reasons over is auditable, not hallucinated. One layer, two audiences: humans get a map of their codebase, agents get a verification loop for cross-file reasoning. Both think more causally, both burn less context, both stop guessing. - **Team:** [Vago Daniel](https://cerebralvalley.ai/u/Dax) - **GitHub:** https://github.com/daxaur/causalist - **Demo video:** https://drive.google.com/file/d/1GNNUeOzgzdrpS3kTTnYkUZ2iZ6P8wayJ/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=246 ### 126. CuraClaw CuraClaw is a personalized AI physician assistant that is designed to cut EMR chart review time by 30 mins per patient, for physicians. It does this by doing deep research on every patient on the list, leveraging standardized FHIR endpoints to interact with patient data inside EMRs to not only create a comprehensive pre-rounding report but to also assist with physician documentation and billing. By eliminating hours of time spent on the EMR, this app aims to solve the physician shortage problem in the US by unlocking their time and thus their capacity to see more patients in the healthcare system. - **Team:** [Sohaib Qadri](https://cerebralvalley.ai/u/sohaib) - **GitHub:** https://github.com/skcadri/curaclaw - **Demo video:** https://www.youtube.com/watch?v=RF8fR4V3EN8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=253 ### 127. Anasan Rai Drop a sales call. Five Claude Opus 4.7 agents qualify, draft, fire, and self-review — in 90 seconds. BridgeFlow Operator is an autonomous five-agent sales floor. Drop a sales call recording or transcript, and five Claude Opus 4.7 agents — Call Analyst, Lead Qualifier, Campaign Architect, Action Executor, and Reflection — qualify the lead, score it HOT/WARM/COLD with a BANT panel, draft a three-touch personalized email sequence, fire real Telegram + Resend + HubSpot actions, generate an importable n8n workflow JSON with a self-correcting loop, and self-review the entire run. Five live integrations, full Supabase persistence, and a voice-first AI consultant ("Jarvis") with Opus 4.7 reasoning plus Haiku 4.5 vision watching the operator's screen in real time. Started two days before the deadline: the first idea was a generic workflow builder, the better idea was the autonomous sales operator we'd actually ship to the real-estate teams we build automation for. Spent $425.21 of $500 in Anthropic hackathon credit across five days — receipts in the README. Architect from V5, build from V1. Live at operator.bridgeflow.agency. - **Team:** [Anasan Rai](https://cerebralvalley.ai/u/builder_anasan) - **GitHub:** https://github.com/anasanrai/bridgeflow-operator - **Demo video:** https://youtu.be/PbhXlRC2o4M - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=256 ### 128. RIDWAN NURUDEEN Every year, billions of people receive adverse government letters they do not fully understand, such as tax assessments, benefit denials, visa refusals, court summonses, and labor disputes. The deadlines are short (14 to 30 days), the legal language is dense, and a paid advisor costs £200 to €400 an hour. Three things happen, and all of them are losing: the poor sign without reading, the middle pay £600 to £1,500 to a paralegal for a one-page reply, and the rest ignore the letter and get garnished, deported, evicted, or sued. An AI legal tool that hallucinates a single citation is worse than ignoring the letter. The user sends a confidently written response that cites a regulation that does not exist and ends up worse than they started. Paperwork is built around that exact failure mode. It is a four-stage pipeline on Claude Opus 4.7: Vision: A model reads any letter (PDF or photo, any language) and extracts the authority, deadline, and key facts. Analysis: Web grounded analysis ranks every realistic response option like appeal, contest, negotiate, or escalate. Drafting: A web grounded draft writes the reply in the source language and proper legal register, refusing to fabricate section numbers it cannot verify. Verification: A second pass verifier independently re-checks every legal citation against primary sources (legislation.gov.uk, gesetze-im-internet.de, eur-lex, court registries) and shows the user a per citation audit table with clickable URLs. When a citation is wrong, one click rewrites the sentence using the source the verifier just found. A counterparty agent then attacks the draft as the issuing authority, parallel researchers gather public evidence, and a reviser rewrites. The user downloads a packet with the response letter, attachments, deadlines, and the citation audit table embedded. This is verified end to end on six public thread anchored cases across the UK, Germany, France, Spain, Brazil, and Canada. The live demo catches its own fabricated citation and self corrects on screen against the actual statute in approximately 70 seconds. - **Team:** [RIDWAN NURUDEEN](https://cerebralvalley.ai/u/Ridwannurudeen) - **GitHub:** https://github.com/Ridwannurudeen/paperwork - **Demo video:** https://youtu.be/xfSDeDS_DAg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=260 ### 129. MetroPrompt MetroPrompt is an agentic pixel art city builder where AI agents collaboratively design and construct a city from your prompts alone. A planner agent and zone sub-agents work together to lay roads, place buildings, and grow neighborhoods in real time. Citizens with Claude-powered brains are spawned into the city, each with unique personalities, needs, and jobs that drive their daily behavior across a 7-day simulation. As AI agents grow more capable, MetroPrompt points toward a future where city planners can rapidly prototype urban layouts, stress-test infrastructure decisions, and observe emergent citizen behavior before a single brick is laid in the real world. - **Team:** [Dhanish Natarajan](https://cerebralvalley.ai/u/danny_nat) - **GitHub:** https://github.com/superstarcoder/MetroPrompt - **Demo video:** https://youtu.be/lT5-Hm6LYc4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=262 ### 130. Roy AI Texas school-board trustees get 100–200 page packets 72 hours before they vote. Most skim. Things get missed — blanket TOMA citations, blanket consent agendas, six-figure financial commitments buried on page 87. The cost is paid by kids. This is the agent I wanted as a trustee. Drop in the PDF. A 3-stage pipeline runs: a generator (Opus 4.7) writes a Brock-format pre-read grounded in a governance doctrine I authored as a sitting trustee; a legal verifier (Sonnet 4.6) fact-checks every TEC/TGC/TAC citation against statutes.capitol.texas.gov and downgrades anything it can't verify; a strategic red-team (Opus 4.7) re-reads each agenda item against the district's own measurable goals and generates the specific data-driven questions a trustee should ask. Output is a Brock-format markdown report with WATCH / RED FLAG / POSITIVE blockquotes, exportable as DOCX. A Q&A chat lets you interrogate the doc; voice is wired through Gemini Live. Tested against my own hand-written pre-reads on real Brock ISD packets. The agent covers the same agenda items in the same structure, with real verified citations, and questions tied to the district's adopted scorecard. The doctrine is the IP — the system extends it; it doesn't replace it. - **Team:** [Toby Farmer](https://cerebralvalley.ai/u/Tobyfarm) - **GitHub:** https://github.com/tobyfarm/TexasGovernanceAgent - **Demo video:** https://drive.google.com/drive/folders/1-vdkk7ey8b1VWq2q8ehSwf5SxzAjM5W7?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=263 ### 131. Immanuel Lam An autonomous agent that re-plans a 64 kWh Sigenergy battery's charge/discharge schedule every 30 minutes against live AEMO wholesale prices, BOM weather forecasts, and Home Assistant sensor data. Every time the plan changes, Opus 4.7 writes a two-sentence rationale explaining what shifted and why. Backtest over 7 days of real household data: agent beat Amber SmartShift's actual dispatch by $5.91/day — by correctly refusing to export when NSW1 prices were negative all week. The win is knowing when not to act. - **Team:** [Immanuel Lam](https://cerebralvalley.ai/u/immanuellam) - **GitHub:** https://github.com/immanuel-lam/sigenergy-nem-arb - **Demo video:** https://www.youtube.com/watch?v=pf-0i2-hWIA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=53 ### 132. Two-Weeks-Team Nobody knows what will get built when a project starts. Specs go stale. Wireframes lie. By demo day half the assumptions were wrong. So we put the picture first. Preview-Driven Development (PDD). Before any spec, before any code, the harness renders the project as nine different mockups in parallel — each a different Opus 4.7 persona pulling your idea in a different direction. You see what could be built. You select one. Preview is all you need. The selection IS the spec. Three cycles run back-to-back: Preview-Driven Development (the gallery is the spec) → Spec-Driven Development (OpenAPI hashed and locked) → Test-Driven Development (5-judge double-gate). Test-Driven Development (TDD) drove code with tests. Spec-Driven Development (SpecDD) drove code with specs. PDD drives them both by deciding what the picture is. The whole run is one autonomous session. Two anchored asks the user knows are coming (H1 selection, H2 freeze). Two adaptive asks the harness fires on its own (Socratic when intent can't be inferred, budget guard at 12× the standard token ceiling). Everything else: auto-decide. Fifty minutes from one line to a working app. Four asks total. Ten of fourteen scenes silent. A conversation with shape. LunchPull, the demo: 44 files, ~3,600 lines, spec conformance 98. Apache-2.0. Single Claude Code plugin. Zero third-party deps. 143 Opus 4.7 personas. Two human clicks. Preview is all you need. - **Team:** [michael chang](https://cerebralvalley.ai/u/sgwannabe) - **GitHub:** https://github.com/Two-Weeks-Team/PreviewForgeForClaudeCode - **Demo video:** https://youtu.be/_xHL8SZqfyI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=96 ### 133. go-ahead River turns chat into a canvas. Every message is a card, every tangent a branch, every turning point a flag. An agent works alongside you as a cartographer — surfacing angles, drawing connections, marking pivots — so the graph of your thinking becomes something you can see, navigate, and build on. Three background micro-agents (skeptic, assumption-checker, expander) offer fresh lenses each turn; tap one to steer the next reply without typing a word. Persistent per-project agent sessions on Opus 4.7 mean it comes back knowing you. - **Team:** [Christopher Pietsch](https://cerebralvalley.ai/u/chrispie) - **GitHub:** https://github.com/cpietsch/river - **Demo video:** https://drive.google.com/file/d/1y1ejEJqnKtjphrnB6s0Gd1ifDBWSc0gN/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=147 ### 134. israNetworks Operator is an AI system that turns marketing data into real business decisions - and executes them. The system integrates four critical layers into a single operational loop: Paid advertising platforms (Meta, Google) Lead sources (forms, CRM systems) Sales outcomes (calls, deals, revenue) External market context (seasonality, demand shifts) How the System Works Operator builds a deterministic feedback loop between campaigns and real revenue. It continuously: Ingests data from ad platforms and CRM systems Matches leads to actual sales outcomes using deterministic logic (IDs, timestamps, deduplication) Evaluates performance based on real business impact - not platform-reported metrics Generates structured actions Executes or escalates those actions based on risk policies This process runs in recurring evaluation cycles (every 4 days), ensuring continuous adaptation to real performance. AI Layer (Opus 4.7 Integration) Operator uses Opus 4.7 as a reasoning layer on top of deterministic data. While core data processing, matching, and evaluation are fully deterministic, Opus 4.7 is used to: Interpret complex performance patterns Generate structured action proposals Provide explanations for decisions Assist in creating new campaign variations This hybrid architecture ensures that: Critical decisions remain grounded in deterministic logic AI is applied where reasoning and flexibility are required Execution Layer (Core Differentiator) Operator is not an analytics tool - it is an execution system. Based on its evaluation, it can: Pause or reduce spend on unprofitable campaigns Reallocate budget toward high-performing segments Prioritize campaigns based on lead quality and actual revenue conversion Generate and suggest new campaign variations All actions are tied to real business outcomes - not proxy metrics like CPL. Controlled Autonomy Operator operates under strict guardrails: Budget changes limited to ±20% per cycle Minimum data thresholds required for action Risk-based execution model: Low → auto-executed Medium → logged with rollback High → requires approval Additional controls: Human-in-the-loop approvals (WhatsApp / dashboard) Full audit logs and explainability Built-in rollback for every action Why This Matters Advertising platforms optimize for signals they can see. Operator optimizes for what actually matters: revenue. Due to tracking limitations, reported conversions often diverge significantly from actual business outcomes. Operator closes that gap by feeding real business data into the decision layer - and acting on it. Scalability The system is designed to operate beyond a single account. It can: Adapt to different industries and business models Learn from cross-account performance patterns Operate across regions with localized context This positions Operator as a scalable execution layer for marketing systems — not just a tool. Core Principle Most systems: → Analyze performance Operator: → Makes decisions → And acts on them - **Team:** [Israel Shitrit](https://cerebralvalley.ai/u/israNetworks) - **GitHub:** https://github.com/isranetworksoffice-dev/LeadOperator - **Demo video:** https://youtu.be/HEuJ007FxjY?si=nP6tltRhVPNThu6e - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=200 ### 135. Overlooked Watcher is a proactive safety agent for Apple Watch and iPhone that takes action when you can't. Before entering a risky context — walking home at night, taking a late ride, going to an ATM — you start a Safety Session with one tap. From that moment, your watch silently monitors heart rate, HRV, motion, and a rolling 30-second audio buffer. When something looks wrong, Watcher doesn't ask you to confirm. It sends the bio-signals, motion features, location, and audio to Claude Opus 4.7, which classifies the situation in seconds: assault, fall, stress, exercise, or false positive. Based on that classification, Claude orchestrates a response — alerting trusted contacts with live location, preserving audio as evidence, escalating only when warranted. Existing solutions (Emergency SOS, panic buttons, fall detection) all assume you can react. The reality of a violent assault is 3-5 seconds before the phone is gone. Watcher inverts the model: the system is already armed, watching, and ready to act on your behalf. Inaction means escalation, not silence. The agent layer is what makes this only possible now. A rules-based system would drown in false positives — a sprint to catch a bus would trigger the same alert as an assault. Claude reasons over the full context: "HR jumped 45 bpm with no warm-up phase, audio captures a male voice demanding belongings, impact spike 3 seconds in, location matches a known high-risk corridor — this is an assault, confidence 0.89." That reasoning is what turns a noisy sensor stream into a trustworthy decision. Built natively for watchOS and iOS using HKWorkoutSession for sustained background execution, Core Motion for fall detection, AVAudioEngine for the rolling buffer, and Claude Opus 4.7 as the decision engine via the Anthropic API. Trusted contacts receive SMS alerts with live location; all events are logged with Claude's reasoning for full auditability. This project was developed with Laura Valeria Mora Parra (lv.morap1@gmail.com) - **Team:** [Juan Sebastian Urrea Lopez](https://cerebralvalley.ai/u/jsurrea) - **GitHub:** https://github.com/jsurrea/Watcher - **Demo video:** https://youtu.be/Ql78EmfmUy0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=255 ### 136. heychoppaah Cognitive Signature IDE — a Claude Code plugin that syncs Claude to your cognitive signature and keeps it aligned as you change. Two functionalities on one pipeline: 1. Voice signature — 7 dimensions of how you direct AI (directive style, compression, reframe patterns, trust signals, idiomatic tells, iteration cadence, texture energy). Extracted from raw user-typed directives across your Claude Code JSONL corpus. 2. Operational signature — 4 dimensions of what you've learned through usage (recurring_decision_templates / recurring_failure_patterns / recurring_tooling_invocations / vocabulary_anchors). Auto-promoted to a permanent signature when n≥2 instances observed. Four onboarding presets (normie / power / team / enterprise) map to three deploy modes (standalone / team-shared / cloud-governed). Hands-off by default; user surfaced only on conflict. Ships 3 dual-function governance subagents — brutus (Opus, adversarial), qa (Haiku, schema/compile/dead-code), historian (Sonnet, drift tracking) — across two substrates (in-session Claude Code plugin + API-level Managed Agents). Same architecture, different persistence. Measurement: 100% auto-scorer accuracy on N=10 blind tests (Claude-as-judge) = +66.7pp over chance, 5-persona simulation grid showing same pipeline produces visibly-distinctsignatures, 5 real governance catches during this build cataloged (QA caught 3 plugin-loader schema violations pre-demo; Brutus caught 3 contaminated blind-test prompts before 30 Opus calls fired; and more). - **Team:** [Choppaaah B](https://cerebralvalley.ai/u/Choppaaah) - **GitHub:** https://github.com/Choppaaahh/cognitive-signature-ide - **Demo video:** https://youtu.be/GvP8P_4Fkl8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=1 ### 137. BHARATH KUMAR P K Compass is a second-brain job search tool I created over 6 days at an Anthropic Claude Code hackathon. I am an Indian MSc student from Bristol, actively seeking visa-sponsorship based roles. And my pain point remains the same, 73 applications sent, received only 4 responses. The challenge is not so much in sending more applications but in targeting and creating warm introductions. Current job search tools help you apply to relevant vacancies according to your CV. Compass works the other way round. It stores all information — skills verified by you, rejections faced, contacts made, artefacts created, etc., and uses 8 agents (Scout, Archivist, Scorer, Visa Checker, Ghost Hunter, Writer, Mirror, Triage) on this brain weekly. The result is a Weekly Compass containing scored job applications, suggestions for warm introductions, draft messages for cold direct messaging, Visa Clock countdown, and skip list. Manually, it is a process that takes around 10-12 hours a week. With the help of Compass, I am now spending 1 hour reviewing this process weekly. There are massive real-world implications of this tool, especially for international students in the UK facing this sponsor-list maze. But for some, it means their rent grows faster than the visa clock. - **Team:** [bharath Kumar](https://cerebralvalley.ai/u/BHARATH9112) - **GitHub:** https://github.com/bharathkumar9112/compass - **Demo video:** https://youtu.be/6LTkvE4E4E4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=54 ### 138. LetItBe gsuda-engine is a provenance-first self-evolving trading loop for Korean equities. It solves a practical problem in trading agent design: generated rules should not be deployed blindly. The system logs simulated trade recommendations with full feature vectors, tracks T+5 outcomes, clusters failed trades, drafts candidate suppression rules, and then validates each rule before Risk Guardian can load it. The project is built from what I know: 23 years operating Korean securities systems and 13 years studying Saju / Yeokhak. Saju is not treated as fortune-telling or price prediction. Instead, original Hanja-based Saju fields are preserved as domain-informed categorical features that must be tested empirically. In the demo, one candidate rule is promoted to skills/active after passing validation, while another is quarantined because it damages too many winning trades. The broader local research warehouse covers Korean equities from 1995 to 2026, with 11M+ enriched rows and 2,700+ stocks. Core principle: Claude drafts. Data validates. Risk Guardian deploys. - **Team:** [HyuckJae Lee](https://cerebralvalley.ai/u/letitbe) - **GitHub:** https://github.com/misolove/gsuda-engine - **Demo video:** https://youtu.be/B8CfS_e32is - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=55 ### 139. Anatolii Project is for business users that want to get data from DB but do not have tech skills. It is always prioritize language explanation versus query data. Main foundations of the project: - For non tech users - Can run only SELECT or DESCRIBE, no destuctive actions allowed - Saves learnings about each table to be better assistant in the future - Always uses limit for queries - One run command - Saves all data locally - Helps user with multi-select questions to get to the right answer - **Team:** [Anatolii Fesiuk](https://cerebralvalley.ai/u/anatolii-asd) - **GitHub:** https://github.com/festoinc/vordbeste - **Demo video:** https://www.loom.com/share/3ebcdf96582e419495f04fec298aa90e - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=57 ### 140. Team HCI I'm an HCI researcher, and I keep running into the same problem: every claim has to be tested with real users. Two to three months of recruitment per study, another six months for the publication cycle, and very little room to discover — halfway through — that something basic about the design was off. Probe is a rehearsal stage for study design, powered by Claude Opus 4.7. You start with a one-sentence premise. Opus splits it into three sub-research-questions — mechanism, intervention, lived experience. Sonnet 4.6 surveys arXiv, ACM, and Semantic Scholar per angle. Opus then drafts three candidate study designs; you pick one, and Probe writes the actual artifacts: implementation plan, IRB memo, validation protocol, survey kit, diary kit. A simulated pilot runs twelve composite personas through the protocol — surfacing failure modes before recruitment. The final stage is an adversarial peer-review panel; Opus 4.7 then audits their disagreement and separates real conflicts from apparent ones. - **Team:** [Bektour](https://cerebralvalley.ai/u/bektour) - **GitHub:** https://github.com/Apolotary/probe-researcher - **Demo video:** https://youtu.be/zB40JrTdV-s - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=97 ### 141. Mycelium Mycelium is an OS built on the same principle as fungi: not one machine, but a mesh. Workstations, edge devices, different CPU architectures — all one living network. Components flow to where they're needed. State replicates everywhere. The capability system is the mesh's immune system. - **Team:** [Antoine Plin](https://cerebralvalley.ai/u/antpln) - **GitHub:** https://github.com/antpln/Mycelium - **Demo video:** https://youtube.com/watch?=... - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=99 ### 142. RED REVIEW RED REVIEW is a forensic audit tool for clinical-AI preprints that produces executable evidence instead of prose critique. Paste an arXiv URL. In under two minutes, it returns a structured red-team report backed by runnable artifacts: a citation whose content does not actually support the paper's claim, a reference domain mismatched between paper and source, a statistical test misused since 2012. Three Claude Code agents run in parallel (Citation Forensics, Guideline Auditor, Statistical Critic) unified through a skills-first architecture with CONSORT-AI and TRIPOD+AI authored as progressive-disclosure item files. The hero capability is Citation Semantic Verification: Opus 4.7 with extended thinking fetches each cited paper, reads it in full, then checks whether the citation actually supports the claim being made. This inverts the "LLMs hallucinate citations" problem, applied to human authors. On a real published paper, it caught a methodology citation that pointed to the test's implementation paper rather than its origin, and mischaracterized the test as a t-test when it is U-statistics-based. AI peer review that produces prose critique manufactures confident wrong answers. RED REVIEW produces evidence. - **Team:** [Tayfur Ertas](https://cerebralvalley.ai/u/Heraldd) - **GitHub:** https://github.com/tayfurertas/red_review - **Demo video:** https://www.youtube.com/watch?v=62zFKwrL9fg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=151 ### 143. The Secretary BoardBreeze Concierge is a 24/7 voice + SMS customer success agent for BoardBreeze (appboardbreeze.com), the AI meeting-minutes app I shipped earlier this year. Built in one week for the Anthropic × Cerebral Valley Global Hackathon, on Claude Opus 4.7. THE PROBLEM BoardBreeze grew faster than I could support it. Subscribers across the US, Canada, England, the Philippines, and Zimbabwe most of them non-technical board clerks, HOA managers, and nonprofit coordinators want to talk to a person on the phone. Some of their questions are easy. Some are even in the FAQ. But I am one person with a full-time civil-service job at City College of San Francisco. I cannot answer every call fast enough. Customer success at that scale needed a real solution. THE SOLUTION One phone number. One Claude Opus 4.7 agent. Five specialist modes inside that single agent, invoked by the agent based on what the caller needs: • Governance Helper — open meetings law, Brown Act, agenda posting rules • Product Expert — BoardBreeze features, pricing, plan comparisons • Tech Support — onboarding and account questions • Sales Closer — answers prospects, books demos • Escalation Handler — texts me directly when a human is genuinely needed That last mode is my dream come true. As an admin assistant, having a system that only bothers you when it truly matters? That's the innovation. Call it yourself: 1-844-786-2076 UNDER THE HOOD • Claude Opus 4.7 — voice and SMS conversation, via the Anthropic Messages API • Claude Haiku 4.5 — fast fact-checker inside the citation-verification tool. Opus thinks, Haiku verifies. • Direct Messages API on voice — I measured ~6 seconds of overhead on the higher-level agent stack, which is a hang-up on a phone call. So the voice channel calls the Messages API directly with sentence-by- sentence streaming. The caller hears the answer beginning while the answer is still being written. • Twilio — voice + SMS • ElevenLabs — text-to-speech in my own cloned voice • Supabase + pgvector + Voyage AI — governance knowledge base with vector search • Citation guardrail — every statute reference is verified to exist AND to support the claim before being read out loud. No hallucinated laws. • Cross-session memory — if Jane texts on Monday and calls on Thursday, the agent remembers her ARCHITECTURE: ONE MANAGED AGENT, TWO TRANSPORT LAYERS The Concierge is a single Claude Managed Agent (CMA) — one agent name, one system prompt, one set of registered custom tools. Following the "one agent + many skills" pattern Anthropic's Michael Cohen described at the hackathon Thursday session. • SMS runs entirely on CMA. Each caller's phone number has its own CMA session. Sessions persist across days and weeks, which is how the agent remembers a caller from a previous conversation — no custom memory layer required. CMA does it. • Voice reuses the same CMA agent definition (system prompt + tools + handlers) but executes the turn loop against the Anthropic Messages API directly, because I measured ~6 seconds of CMA event-stream overhead per turn — unworkable inside Twilio's 15-second webhook ceiling. • Four custom tools registered with the CMA agent: search_governance_kb, search_product_kb, verify_citation, escalate_to_grace. • The agent evolves in place via client.beta.agents.update() — added a tool mid-hackathon without resetting any existing caller's session memory. PRODUCTION DEPLOYMENT Pivoted from AWS App Runner to AWS ECS Fargate mid-hackathon. Docker containerization, Amazon ECR, AWS Secrets Manager for 11 API keys, two least-privilege IAM roles, Application Load Balancer with Amazon-issued TLS certificate, custom domain at concierge.appboardbreeze.com via Vercel DNS, CloudWatch logs, deployment circuit breaker with auto-rollback. Live at: https://concierge.appboardbreeze.com Built with Claude Code as my development partner. - **Team:** [Grace Esteban](https://cerebralvalley.ai/u/gracesteban) - **GitHub:** https://github.com/mgesteban/concierge - **Demo video:** https://youtu.be/znKrItMsI-I - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=154 ### 144. OpusGraph Citation networks treat every paper as an atomic node — you can see which papers cite which, but not what they actually cite at claim level. Two papers citing Zhao 2014 might be citing two different ideas in that paper. Opus Graph uses Claude Opus 4.7's native multimodal PDF reading (no pre-parsing — text, figures, equations, tables fed straight in) to decompose each paper into typed claims (methodological, empirical, theoretical, scoping) and labelled inter-claim relations (supports, depends-on, extends, contradicts). Twenty urban-climate papers become one queryable claim graph: 365 claims, 407 relations, 831 edges across 385 nodes. The headline query walks lineage between papers at claim level. For the anchor pair Sun 2017 → Li 2024, it finds a shared theoretical grounding in Zhao et al. 2014 — not by parsing bibliographies but by traversing the corpus-level claim graph. Three delivery surfaces — a Claude Code skill plugin, an MCP server, and a CLI — all wrap the same Python library. Schema validated against a hand-authored exemplar at 80% lexical recovery; zero out-of-vocabulary labels across the corpus. - **Team:** [Ting SUN](https://cerebralvalley.ai/u/sunt05) - **GitHub:** https://github.com/sunt05/opus-graph - **Demo video:** https://youtu.be/bFkC5enLIfU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=159 ### 145. Fifty Eleven AI Solving complex problems and using the frontier models to long research tasks. Not just research or deep search but the ability for it to be done over time while the user is away. If new information is found while the agent is researching it follows it's path to help make its decision. Its not just deep research here is a 2 page paper, it's research over time and allowing new facts to be introduced while that same string of research is ongoing. I wanted to challenge how long I could make the model think. - **Team:** [Ian Walmsley](https://cerebralvalley.ai/u/packetloss404) - **GitHub:** https://github.com/packetloss404/vellum - **Demo video:** https://www.youtube.com/watch?v=tB2kS0MhOb0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=162 ### 146. One Piece of Data An interactive, LLM-extracted relationship graph for the entire One Piece manga: every character, crew, organization, saga, and arc, connected by 19,320 typed edges — fought, defeated_by, ally_of, enemy_of, member_of_crew, family_of, mentor_of, and more. Existing One Piece databases store facts (bounties, appearances). They don't capture relationships — who fought whom, who's allied with whom, who's whose family — because those live in 13,000+ unstructured wiki sections. We extracted them in a single 21-minute Claude Sonnet 4.6 run for ~$91, with a verbatim evidence quote attached to every edge so the graph is auditable, not a black box. The pipeline reads One Piece Fandom wiki sections, runs them through Claude with a constrained 12-relation vocabulary and a strict "quote ≥5 words from the source or skip the triple" rule, deduplicates the 43,814 raw triples down to 19,320 unique edges, and ships them to Supabase. A React + vis-network single-page app then renders the graph live: pick any entity to focus on, expand 1-4 hops at a time, filter by confidence and relation type, and click any edge to read the exact wiki sentence that justified it. A sortable edge table below the graph lets you sort 19K rows by confidence or chapter and click straight through to that edge in the visualization. Three things make it work in production: (1) hash-based change detection so weekly re-runs after each new chapter cost ~$2 instead of $91; (2) Anthropic's structured-output JSON schema, which gave us zero parse failures across 13,119 calls; and (3) BFS-based subgraph rendering with an edge cap, so a graph with 19K edges stays interactive in the browser at ~1.5 MB gzipped. The codebase is fully reproducible from make commands and split across two AGPL-3.0 repos: a Python data pipeline and a React web app. Alongside the Story Graph we also shipped a few smaller features in the same week: a new Character Occupations page that surfaces every role a character has held (1,236 characters, 2,038 role assignments); a Chapter Release Predictor that forecasts the next 5 chapter dates from the live Shonen Jump issue schedule; devil-fruit and haki badges added to the character comparison tool and the main characters table with a column-visibility toggle and persistent table state; and shareable result images (feed and story formats) for both the Guess the Character and Who Am I? games, plus a tiered difficulty rework of the Who Am I? quiz with a recalibrated speed-based ranking system. URLs: - https://onepieceofdata.com: the web application - https://onepieceofdata.com/#/analytics/story-graph : the story graph - https://github.com/ismailsunni/onepieceofdata : the brain, the parser, the data generator repository - https://github.com/ismailsunni/onepieceofdata-react : the web app repository - **Team:** [Ismail Sunni](https://cerebralvalley.ai/u/ismailsunni) - **GitHub:** https://github.com/ismailsunni/onepieceofdata - **Demo video:** https://youtu.be/MACaCSYItVs - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=203 ### 147. BloodAI BloodAI is an end-to-end clinical triage system that bridges the gap between a confused patient and an overworked doctor. PROBLEM: In Poland, doctors have 7 minutes per patient. 47% of patients don't understand their lab results. Wait time for a specialist on public insurance is 3-18 months. SOLUTION: Photograph your lab report. In 10 seconds: - Opus 4.7 Vision extracts 9 lab parameters - Custom BERT model (18MB, trained from scratch on 408,214 Synthea+MIMIC sequences) performs 8-class triage with calibrated probabilities (ECE 0.012) - 52 vocab-aware adaptive questions empirically shift predictions by up to 37 percentage points - Opus 4.7 generates patient-mode (plain language) and clinical-mode (structured assessment) explanations - 37 personalized follow-up tests recommended with 14 conditional rules - Real doctor finder integrates live NFZ public health in Poland API + Google Places for private clinics - Trend analysis with 6 cross-parameter comorbidity detectors (CKD progression, pancytopenia, etc.) with R^2 up to 0.99 13 live endpoints. Zero mocks. Built solo in 36 hours. - **Team:** [Antoni Pater](https://cerebralvalley.ai/u/Bobi) - **GitHub:** https://github.com/antoniopater/BloodAI-Hackathon-Claude - **Demo video:** https://youtu.be/gD1mwlyJXwI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=208 ### 148. BD **Cross-turn reasoning, rendered live. Both sides leave with something.** Lacunex runs goal-directed adaptive interviews where structured insight is produced *during* the conversation, not in an overnight report. A Host (subject-matter expert) defines objectives; the platform conducts every turn live, surfaces cross-turn observations the moment they cohere, and at session close hands the participant a reflective takeaway worth keeping. Built by a systems analyst, ten years in. Lacunex closes two gaps real-world interviewing never closes: beginners — analysts, PMs, anyone for whom interviewing is *part* of the job — don't yet know how to run a round well; experienced researchers do, but lack the time to interview enough people for long enough. A four-call Opus 4.7 architecture runs every turn: - **Conductor** — decides the next move from session state and renders the interviewer's turn. - **Meta-Noticing** — observation-only; spots contradictions, hedging, and implied-not-said across turns. Every notice must cite ≥2 distinct transcript anchors (enforced in code, not just in prompt). - **Extraction** (Haiku 4.5) — schema-bound; fills the host's live insight dashboard turn-by-turn. - **Takeaway Synthesis** — a reflective artifact for the participant, including a "what you already have that is relevant" section. Mapped against Anthropic's *Building Effective Agents* (Schluntz & Zhang, Dec 2024): **three of the five patterns layer on a single turn — Orchestrator-Workers, Parallelization, Routing.** Three domain briefs ship — **Founder Investment Evaluation**, **Post-Incident Witness**, **Civic Consultation** — plus **Brief Designer**, where the platform interviews the host to author a *new* brief using the same four calls. Recursive dog-food. **Audiences exercised during build:** VC and angel investors evaluating founders; city officials running civic consultations; SREs, safety officers and clinical QI managers conducting post-incident interviews; clinicians eliciting patient values; tacit-knowledge capture from retiring experts; consumer-insights teams; qualitative researchers; managers preparing for high-stakes conversations. Empirical scale, on a **simulated 11-resident cohort** run against the Civic Consultation brief on the real platform: **303 turns, 54 deployed `◆` cross-turn observations from 243 considered, 12 cohort patterns, 6 routing recommendations.** Cost with prompt caching: **~$1–2 per session**, under $5 for the cross-cohort aggregate — putting a 100-resident consultation at ~$150 end-to-end. **Not an overnight research moderator** (Outset, Listen Labs, Strella ship next-day reports). **Not a transcript analyser** (Dovetail, Condens ingest transcripts that already exist). **Not "Claude with a long system prompt"** — a single chatbot can't enforce cross-turn reasoning the way four calls with code-enforced turn anchors can. - **Team:** [Benjamin Dysin](https://cerebralvalley.ai/u/benjdy) - **GitHub:** https://github.com/Attius-Digital-Art/lacunex - **Demo video:** https://www.youtube.com/watch?v=UyxJRWqi-4I - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=213 ### 149. Tio Cumbana Tio Cumbana is a digital agronomist for smallholder farmers — 500 million globally, 3.2 million households in Mozambique alone, feeding 2.5 billion people but invisible to agronomy. Most AI waits to be asked. Tio Cumbana doesn't. A long-running Claude Managed Agent watches each parcel, cross-references the farmer's photos, weather (via an MCP server), phenological stage, and Zimpeto market prices, and decides autonomously when to interrupt — sending a voice note in Portuguese with Changana code-switching. Two farmers asking the same question receive different answers because they are two different people with two different parcels and two different histories. I am the user. I run a 54-row farm in Maluana with my mother. The nearest agronomist is 40 km away. Today I came back to find a pest destroying our crop. This is the use case. The same Memory layer also gives supermarkets a live harvest-pipeline view, lowering waste and enabling pre-purchase contracts. One agent, every node of the value chain — only viable now, on Opus 4.7 + Managed Agents, at cents per farmer per day. - **Team:** [Elton Laice](https://cerebralvalley.ai/u/Elton) - **GitHub:** https://github.com/eltonlaice/tio-cumbana - **Demo video:** https://www.loom.com/share/4658a6db42994bc39c212d1c2759aea9 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=218 ### 150. Gabriel Solomon (Offset Works) There's so much happening in any city on any given weekend — protests, town halls, mutual aid runs, free concerts, volunteer cleanups. And most people miss most of it. Not because they don't care. Because the information is everywhere and nowhere: split across Instagram, Telegram, Eventbrite, city websites, neighborhood listservs, and handmade flyers on telephone poles. Convoca is my attempt to fix that. It's an open-source platform that uses seven Claude agents working together to find, read, deduplicate, and surface community events in real time — across sources no single person has time to monitor. The agents parse what you're actually asking for, pull from dozens of public feeds, extract structured events from bilingual flyers using vision, and merge duplicates across platforms — showing exactly why two posts are the same event. Every result comes with a plain-language reason for why it was recommended to you. But it's not just search. You can save events, have the agent build a Saturday itinerary around your causes, and get email reminders before things you care about. If you're on the ground and see something, you can submit a real-time safety flag — police presence, route change, medical aid station — and it goes through an AI review and onto the map in seconds. The whole thing streams its reasoning live. You watch the agents work. Every decision has a visible explanation. I wanted people to feel like the agent is working for them — not just returning results, but showing its thinking so they can trust it. Three rules I didn't bend on: public sources only, no user profiling, and the community is the main input channel. One tap to submit a flyer. The vision agent does the rest. It works for any city. Add your sources, point it at your community, ship it. The infrastructure is there. The organizing doesn't have to start from zero. - **Team:** [Gabriel Solomon](https://cerebralvalley.ai/u/Offsetworks) - **GitHub:** https://github.com/0xgasc/convoca - **Demo video:** https://drive.google.com/file/d/1Wrtnw57dG1iEUBT5TkKr8V0lvhl8aDDF/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=221 ### 151. Bibek RoadWatch is an autonomous multi-agent pavement surveyor that grades road condition at scale using Mapillary street-level imagery — no custom CV model, no windshield survey crew, no $250k engineering contract. A 3-tier Opus 4.7 hierarchy works like a field crew: Street Captain (1×) plans the inspection route, dispatches survey teams, synthesizes a corridor narrative Point Surveyors (up to 3×, parallel) investigate each survey point across all available years — cross-witnessing temporal change Year Investigators (up to 2× per Surveyor, parallel) examine the same spot from different epochs, sharing claims on a blackboard so siblings can anchor their yaw angles Peak concurrency: 6 simultaneous Opus 4.7 calls. The discipline gate is deterministic — it inspects the visit log and refuses a grade() call if multi-year coverage exists but only one year was examined. A SHA-256 collision detector identifies when "multi-year" listings are actually the same image with backdated metadata — surfacing a data integrity problem invisible to any pipeline that trusts timestamps. The web UI lets a public-works engineer draw a polygon, select streets, watch the agent hierarchy work in real time — cursors walk the block, evidence cards stream in, the map recolors as grades land. One click, ~$5, any street in the world with Mapillary coverage. Validated: Spring St LA, Chestnut Ventura, SoHo Broadway. Calibrated against LA's published 69,282-segment PCI dataset. - **Team:** [Bibek Parajuli](https://cerebralvalley.ai/u/Bibek624) - **GitHub:** https://github.com/bibek624/roadwatch - **Demo video:** https://www.youtube.com/watch?v=XVfxFR1nMwg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=257 ### 152. DEFAI_HQ DEFAI ComplianceOS is a cross-border crypto-asset compliance reasoning layer. It returns pre-transaction PASS / FLAG / BLOCK decisions across VARA (UAE), MAS (Singapore), FCA (UK), and FATF in a single API call, with verbatim regulatory citations grounded in the source rulebook PDFs. Existing tools (Chainalysis, Elliptic, TRM Labs) flag risky wallets and patterns. None of them tell you why, in the language of the regulator who'd actually audit you. ComplianceOS does. Built on Claude Opus 4.7 with pgvector RAG over the actual VARA, MAS, FCA, and FATF PDFs, plus a content-hashed audit trail (SHA-256 prompt hash + regulatory snapshot) so any decision is reconstructable months later. OFAC SDN wallet screening runs deterministically before Claude is called. Status: 5-day proof-of-concept built for the Anthropic Built with Opus 4.7 Hackathon. Chapter 1 of three — the compliance-reasoning foundation. Chapters 2 (agent identity via ERC-8004 + x402) and 3 (B2B payment rails on stablecoin). - **Team:** [Samuel uzoamaka](https://cerebralvalley.ai/u/SIgma-Dev) - **GitHub:** https://github.com/Sigma-Dev001/DEFAI_ComplianceOS - **Demo video:** https://youtu.be/TZbuB_EqJLA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=2 ### 153. Postmortem Postmortem is a decision-archaeology agent for any GitHub repo. Point it at a codebase and it reads every PR review, every rejected alternative, every architectural debate — and builds a queryable decision ledger. Ask why the code is the way it is and get a cited answer back in seconds, with every claim traceable to the exact reviewer quote that decided it. Five interaction modes ship: 1. Ask — natural-language query over the ledger; answers stream with live adaptive-thinking tokens and a self-check pass that verifies every inline citation before the answer is trusted. 2. Impact Ripple — BFS across the decision graph to trace what breaks when an assumption moves. 3. Open Decision — the full rationale + rejected alternatives for any single PR, cited. 4. Ghost Interview — pick a maintainer; Opus speaks in their register using only their own verbatim quotes, with every paraphrase openly disclosed. 5. Conflict Finder — scans the whole ledger for pairs of decisions that quietly contradict each other, ranked by severity. Shipped with six hero ledgers (hono, zustand, next.js, shadcn-ui, supabase, and the Postmortem repo itself) totaling 155 decisions, 1,876 citations, and 410 rejected alternatives — a queryable public-domain artifact. Also ships as a Claude Code MCP server so all five modes are one claude mcp add postmortem away. Live: postmortem-mauve.vercel.app · Video: youtu.be/9nbIVa_jnsU · 100% citation-grounded. - **Team:** [Rahil Singhi](https://cerebralvalley.ai/u/Rahilsinghi) - **GitHub:** https://github.com/rahilsinghi/postmortem - **Demo video:** https://youtu.be/9nbIVa_jnsU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=56 ### 154. Dima Even top ML PhDs cap out at a 41.4% replication score after 48 hours of dedicated effort per paper (PaperBench, OpenAI, 2025). Full replication is too slow and too hard, even for experts — which is exactly why static diagnosis has to be the first line of defense. The bottleneck isn't compute, it's diagnosis. The bug — an augmentation leak, a train-test overlap, a metric mismatch, a missing `model.eval()` — was almost always a static read away. RunItBack does the static read. Point it at a paper, a repo, and some data; four Opus 4.7 agents (Paper Analyst, Code & Data Auditor, Validator, Reviewer) running on Claude Managed Agents return a diagnostic report with a verdict, claim-by-claim verification, severity-ranked findings, and unified-diff fixes — in minutes, before any GPU starts. Opus 4.7 ingests PDFs natively (tables, figures, equations — no OCR, no `pdftotext`), carries the full reproducibility-failure taxonomy across 60-turn tool-using sessions, and enforces a ≥ 2-agent cross-check rule that surfaces genuine disagreements instead of hallucinating consensus. Managed Agents provides the sandbox, tools, sessions, and streaming on day one — RunItBack adds only the brain. - **Team:** [Dmitry Golovchits](https://cerebralvalley.ai/u/golovchits) - **GitHub:** https://github.com/golovchits/RunItBack - **Demo video:** https://drive.google.com/file/d/1c0yJFpPJjAiKCjmJxmjYwY-jLp1eLEur/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=98 ### 155. cooleel Breakpoint — a debugger for AI agents. Every tool call the agent makes is checkpointed inside a Tensorlake sandbox (filesystem + memory + processes). Opus 4.7 reads the full trajectory — reasoning, tool I/O, filesystem deltas — and points at the causal tool call that broke the run. You can rewind to any checkpoint, scrub through the timeline, live-shell into past sandbox state, or fork a fresh agent from any step with a corrective prompt. Drop-in for Claude Agent SDK; framework-agnostic adapter (bash / edit_file / view) works with OpenAI Agents SDK and any harness that registers Python functions as tools. Problem it solves Agent failures are causal, not local. A model wipes a database on turn 3, looks fine, succeeds on turns 4–9, then fails an integration test on turn 10 with a confusing error. Existing tracers (LangSmith, Langfuse) show you the crash site — turn 10. The actual fix lives at turn 3. That's a git bisect problem on a single agent run, and today there's no good way to do it: - Logs and traces show messages, not state. You can read what the agent said, but not what the filesystem, processes, or database looked like when it said it. The real evidence — a stack trace in /tmp/server.log, two missing rows in todos.db, a half-written config — is gone by the time the run ends. - Re-running doesn't reproduce. Agents are non-deterministic; reproducing a specific failure path can take many tries. - There's nowhere to fork from. Even if you spot the bad step, you can't restart the agent from that exact state with corrective guidance — you have to re-run from scratch and hope it makes the same mistakes up to that point. Breakpoint fixes all three: every tool call is a rewindable checkpoint of full sandbox state, Opus 4.7 does the bisect for you (and can cat log files in any past snapshot to confirm), and any checkpoint is a fork point — restart from there with a fix, watch it run live, get an automatic verdict on whether the fork actually succeeded. - **Team:** [Shanshan Wang](https://cerebralvalley.ai/u/cooleel) - **GitHub:** https://github.com/cooleel/Breakpoint - **Demo video:** https://youtu.be/H2bTq-v0KzI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=152 ### 156. Alexis Chapellier Every year, 50 million tons of electronics become waste — not because they can't be repaired, but because board-level expertise is locked in a few hands. Millions of independent repair technicians, from Chennai to Shenzhen to my own workshop in France, fix screens and batteries — but stop at the motherboard. Wrench Board unlocks that last mile. An agent-native diagnostic workbench powered by Claude Opus 4.7: type a device label, and four Claude agents — Scout, Registry, Writers, Auditor — produce a verified repair pack in two minutes. Upload the schematic PDF, and Claude vision compiles it into a queryable electrical graph. Then the agent takes over: it interrogates the graph instead of guessing, pilots the boardview visually, highlights components, draws boot sequences phase by phase, proposes auditable diagnostic protocols with measurement widgets, cites its sources, and never fabricates a reference designator. Across sessions, every finding is mirrored to a per-device Managed Agents memory store for cross-session continuity. Built and dogfooded in my own repair shop. Right-to-repair becomes real when one technician with a multimeter can do what only an OEM service center could do yesterday. - **Placement:** 2nd Place - **Team:** [Alexis Chapellier](https://cerebralvalley.ai/u/Jnkz) - **GitHub:** https://github.com/Junkz3/wrench-board - **Demo video:** https://youtu.be/OZ2D_p82z6w - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=204 ### 157. Holo Director AI has made it easy to create, but it has also pushed us to create outside our skill set. An engineer might struggle to judge whether generated art is actually good, or a lawyer might not know whether the music they made will connect with an audience. That is why I built HoloDirector: a creative simulation engine powered by Opus and Claude sub-agents. You give a rough idea and data. HoloDirector builds a simulated target audience that reacts to your work based on your input, influences each other, and shows what feels exciting, confusing, generic, or worth improving before you publish. For the demo, I used HoloDirector to create and judge something outside my own skill set: a VTuber idol group. However as Holo Director takes video input to analyze your work, the potential usage is pretty wide range from evaluating a recorded website, gameplay experience, movies and many more. In the future I believe simulation will be used by creators to help create more personalized and better content/app for their end users. Can't wait for the model get better. - **Team:** [Jason Sujaya](https://cerebralvalley.ai/u/JasonSujaya) - **GitHub:** https://github.com/JasonSujaya/HoloDirector - **Demo video:** https://youtu.be/xQqgArNUX5k - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=209 ### 158. SFWIZ SFWIZ is a terminal-based agent harness for Salesforce development, inspired by Claude Code but purpose-built for the platform. When you work on Salesforce day to day, you end up juggling org auth, metadata deploys, anonymous Apex runs, SOQL probes, and release notes that shift every season. Generic LLM agents don't really know the platform, and they forget the moment a new release ships. SFWIZ tries to fix both problems at once. Under the hood there's a single orchestrator running on the Anthropic SDK with streaming and prompt caching. It routes each turn to one of six persona subagents through the Claude Agent SDK: org-admin, designer, developer, deploy-manager, reviewer, and qa. Each persona has its own strict tool-scope, so the reviewer is read-only and only the deploy-manager can actually push metadata. Anything destructive (deploys, scratch creation, permset assignment) goes through a mandatory ask_user gate that blocks the agent until you confirm. For lifecycle work it shells out to the sf CLI rather than reimplementing what's already there, and for runtime queries it uses jsforce. A background Bun Worker quietly scrapes the Apex reference, the LWC guide, and the last few release seasons into a local qmd RAG store, so the agent gets smarter about Salesforce over time instead of staying frozen at its training cutoff. The whole thing ships as a single compiled binary. - **Team:** [Alfian Busyro](https://cerebralvalley.ai/u/arufian) - **GitHub:** https://github.com/arufian/sfwiz - **Demo video:** https://www.youtube.com/watch?v=PNuuCcX8wpM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=220 ### 159. HepVar Five years ago my wife was misdiagnosed with an autoimmune liver disease and put on lifelong immunosuppressants that didn't fully work. After four specialists and years of side effects, a fifth one finally suspected a rare genetic liver disease and ordered a genetic test. It found a mutation in a liver gene — but the lab couldn't tell us whether it caused her disease. Verdict: *"we don't know." That answer came back not because the mutation was harmless, but because she's Brazilian, and the world's genetics databases are built mostly from European DNA. My wife is a geneticist. I'm a molecular biologist. We spent weeks rebuilding the evidence by hand, including a study where researchers had introduced her exact mutation into cultured cells and shown that it disrupts the protein's function. She switched medications and was healthy in a month. HepVar is that workflow, automated. It pulls a dozen genetic, structural, and literature sources in parallel and hands the package to Claude Opus 4.7 with extended thinking. Opus weighs conflicting evidence, names the specific next experiment that would resolve the uncertainty, and unprompted flags clinical safety issues. Three modes (technical / clinician / patient) re-frame the same report without changing a single number. Open source, MIT. - **Team:** [Ariel Maia](https://cerebralvalley.ai/u/bionisch) - **GitHub:** https://github.com/arielmmaia/hepvar - **Demo video:** https://drive.google.com/drive/folders/1mD021E_KC7UhO-83JQ5Ag8_k8aXlrfhA?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=270 ### 160. Mobius You drop a physics paper as a PDF. You get back an interactive simulator that runs the actual physics from the paper, validated against the paper's own figures. The visual style borrows from Bartosz Ciechanowski's interactive explainers. There's one Opus 4.7 model in the middle, acting as a conductor. It drives nine specialist skills along an explicit DAG over seven phases. In Phase A, three parsers run in parallel: paper-parser, physics-interpreter, claim-extractor. They fan into a typed SimSpec. From there the pipeline fans back out into Python primitive generation, sandboxed execution inside Anthropic's Managed Agents Python runtime (numpy, scipy, sympy), and an interactive scene wired to those primitives over SSE. When you move a slider in the scene, the solver re-runs in the sandbox. It's not animation, it's the real solution being recomputed. Every output goes through a parallel science-integrity critic that runs eight orthogonal checks. Pre-render: units, CFL stability, conservation residual, claim-match, eligibility. Post-render: perceptual figure-diff using Opus 4.7 vision, narration-bijection, multimodal triangulation. The problem I'm trying to solve: physics papers ship as static PDFs, but the equations they describe are dynamic interactive systems. Reproducing a paper's figures normally takes a domain expert, a numerical stack, and a couple of weeks. Mobius makes the paper *be* the simulator. About 25 minutes per run, on any PDF, with the actual physics, perceptually checked against the source. - **Placement:** Finalist - **Team:** [David Hayot](https://cerebralvalley.ai/u/leventilo) - **GitHub:** https://github.com/leventilo/mobius - **Demo video:** https://youtu.be/ZODrihiz96Y - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=273 ### 161. Thom Pham Modern AI systems can produce outputs that are perfectly structured—and completely wrong. This demo highlights a failure mode where models prioritize schema compliance over input fidelity. When a constraint is violated—such as requesting twelve tickets when the schema allows a maximum of four—the model often adjusts the input to fit, rather than rejecting it. The result is valid JSON that appears correct but is not grounded in the original request. This is not a reasoning failure. It is a constraint handling failure. The demo was originally designed as a retry loop, but I shifted toward robustness under messy input once Opus 4.7 handled most clean cases in a single pass. ValidJson uses a minimal stack: five rules, one system prompt, and a JSON Schema validator. Instead of always producing an answer, it enforces refusal when constraints cannot be satisfied. Valid JSON guarantees structure—not truth. Refuse > Invent. - **Team:** [thomas Pham](https://cerebralvalley.ai/u/Guardiania) - **GitHub:** https://github.com/GuardianAI1/validjson - **Demo video:** https://www.youtube.com/watch?v=8wLu9GOpzlI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=3 ### 162. Maieutic Maieutic flips the role an LLM usually plays in learning to program. Instead of handing the student working code, it guides them to think first about what the program should do — the specification — before they can write a single line of code. When the code is done, the student has to explain the differences between what they said they'd do and what they actually wrote. This is built on a bet about how programming is changing. Writing code from scratch matters less than it used to; specifying behavior precisely, reading code critically, and noticing the gap between intent and output matter more. Those are the skills Maieutic trains. Along the way it surfaces a learning signal that's normally invisible to an instructor: not just whether the code passes the tests, but whether the student can explain why their code does what it does, where they drifted from their own spec, and whether they can say why. That's what a teacher needs to see individual students develop — and it's nearly impossible to capture by hand in a lab with forty students. - **Placement:** 3rd Place - **Team:** [Paula Vasquez-Henriquez](https://cerebralvalley.ai/u/pauvasquezh) - **GitHub:** https://github.com/pauvasquezh/maieutic - **Demo video:** https://www.youtube.com/watch?v=IJ9FyX2xwWA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=58 ### 163. 3up Technology [The Problem We Solved] The biggest hurdle in building industrial robot systems is that experienced engineers must read through massive manuals (hundreds of pages of communication protocols and specifications) from different manufacturers to manually set up and configure connections one by one. This process typically takes weeks, placing a heavy burden on System Integrators (SIers). [Project Details] In this project, we used Claude Code (Opus 4.7) to automate this manual-heavy, tedious integration work to dramatically streamline SIers' operations. We integrated five entirely different devices that usually require reading separate manuals: a DENSO 6-axis robot, an IAI electric cylinder, a Basler industrial camera, a USB camera, and a PS5 controller. We successfully automated the entire setup—from IP configuration and communication checks to building a basic remote control system—in just 4.5 hours. This demonstrates a future where SIers are freed from the unproductive task of deciphering manuals and can focus on more advanced system design. - **Team:** [Norihide Mikami](https://cerebralvalley.ai/u/3up) - **GitHub:** https://github.com/mikami-robot/denso-robot-control - **Demo video:** https://youtu.be/-cgO3RlOce4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=100 ### 164. Haruto Kimura TriageGuard is an autonomous validator for vulnerability reports. Drop in a report + PoC + claimed affected code; it returns a Signal-vs-Slop verdict with auditable reasoning, end to end and unattended. The bug bounty industry is collapsing under AI-generated low-quality reports. HackerOne paused the Internet Bug Bounty program on 2026-03-27 after 13 years; curl ended its program in January; valid-submission rates at major OSS programs fell from ~15% to under 5%. The bottleneck has shifted from **discovery** to **validation** — and that is the bottleneck TriageGuard attacks. The orchestrator (Python, Claude Agent SDK) fans out four parallel Opus 4.7 sub-agents at `xhigh` effort: - **Reproducibility** — clones the target, builds with ASan in a Docker sandbox, runs the PoC, parses sanitizer output. - **Root Cause** — opens the claimed source and checks the data-flow claim. - **Duplicate** — queries NVD via a custom SDK MCP tool. - **Hallucination** — cites or rejects every concrete technical reference (function names, line numbers, CVE IDs). Each sub-agent writes a strict JSON artifact to `findings/{report_id}/`. A **deterministic synthesizer** (no LLM) maps the four artifacts to a 0–100 Signal Score using the rubric in `signal-score-rubric/SKILL.md`. The verdict is auditable, reproducible, and cheap. Haiku 4.5 runs a preflight digest before fan-out (~$0.0015, ~1.5s) that extracts the report's claimed bug class, locations, and evidence — surfaced in the CLI and the synthesizer narrative as a fast sanity check. The web UI streams the same artifacts via SSE. Primary target: **wolfSSL** (~5B devices). The eval set includes two of my own published wolfSSL CVEs (CVE-2026-3849, CVE-2026-2646), plus two public curl slop reports and one live-generated slop sample. Built by a bug-bounty researcher with 9 published CVEs across wolfSSL, Mozilla NSS, and PowerDNS. *"I helped create this crisis. TriageGuard is my contribution to solving it."* - **Team:** [Haruto Kimura](https://cerebralvalley.ai/u/hkimura) - **GitHub:** https://github.com/HarutoKimura/triageguard - **Demo video:** https://youtube.com/shorts/jqrxefYqbO4?feature=share - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=153 ### 165. CEOM NISP is a 5-layer neuro-symbolic pipeline that reads a foundation model's predicted brain response to autonomous systems videos and produces a formally certified anomaly verdict: slow enough for a Lean 4 proof, fast enough for real-time alerting. - **Team:** [Hans Magix](https://cerebralvalley.ai/u/abs_jaded) - **GitHub:** https://github.com/absjaded/nisp-v0.2.0 - **Demo video:** https://youtu.be/gi1fEpLfYn0?si=F1qmppB9VqEzEYnI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=205 ### 166. ZLDA Labs OpsAdvisor is an operations diagnosis and improvement tool that runs the same Lean Six Sigma workflow I run manually today, compressed into a single structured pass. My background is operations, systems, analytics, and process improvement across healthcare, legal, and construction. In my current role I'm building structure inside a fast-growing physician group where complexity lives across SOPs, handbooks, handoffs, and reporting. The hard part isn't storing that information. It's turning it into principled insight and clear next actions, and that is the gap OpsAdvisor is built to close. For this MVP, OpsAdvisor focuses on the Lead Intake and Conversion Bottleneck, a problem almost every services business has and almost none diagnose rigorously. It takes a business context brief, a process note, and workflow data, then runs a structured Define, Measure, Analyze, Improve, and Control flow to identify where the process is leaking, what is driving the loss, which fix should happen first, and how to monitor the change so it does not decay. The output is not a chat transcript and not a dashboard. It is an executive summary, baseline metrics, root cause analysis, a ranked recommended fix, a workflow rule update, and a control package with monitoring guidance and alert thresholds. As a concrete example, a run might surface that Stage 3 to Stage 4 conversion drops to 18 percent driven by SLA breaches on intake callbacks, and recommend a 4 hour callback rule paired with a daily breach alert wired into the control plan. Three things I would want judges to weigh. First, the output is deterministic in shape and validated against a TypeScript contract before it leaves the API, so the model cannot quietly change the report structure between runs. Second, nothing is persisted. Analysis runs in session and any database credentials supplied for a one time pull are never stored or logged, which is a deliberate choice for the kind of operational data this tool is meant to handle. Third, this is genuinely the diagnostic I do by hand for real operators today, not a hypothetical use case dressed up for a demo - **Team:** [David Tanis](https://cerebralvalley.ai/u/Tanis226) - **GitHub:** https://github.com/tanis226-hue/operator-agent - **Demo video:** https://www.loom.com/share/e1fbff0ad5a24d56b30d9ea4db564db3 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=215 ### 167. SemSwitch CrowdPatch is a closed-loop bug fix economy. Developers upload their apps, the users file bug reports and earn credits, and uploaders spend credits to spawn Claude Managed Agents (Opus 4.6) that diagnose, fix, validate against the repo's own test suite, and open real GitHub PRs. Every other AI coding tool requires a developer in the loop. CrowdPatch inverts the model: the crowd becomes QA, the agent becomes the engineer, the developer becomes the curator. Every successful run is a durable, reviewable, mergeable PR, not ephemeral output. Here's what we got: the agent + workers. A Cloudflare Worker (Hono) hands fix-jobs to a Durable Object that owns one Managed Agent session per job, multiplexes its event stream to concurrent SSE browser subscribers, persists semantic events to per-instance SQLite for replay-on-reconnect, and writes a per-run R2 evidence bundle. Meanwhile, a second Vault + GitHub MCP path lets user-supplied PATs author PR commits, solving the read-only-token limitation of the default github_repository.authorization_token attachment. Evidence: 14+ real [CrowdPatch] PRs landed on a public test repo across the build week. PR #14 is the submission-day headline. PR #13 is a resilience receipt - a CF Durable Object died at the +5min CPU limit, the agent kept working and pushed its branch anyway, and the recover endpoint adopted the orphan branch and opened the PR. Open source under MIT. The agent + worker spine is the reusable core; the website is optional. But I felt like if I didn't start the rollout, someone else might have done it. So, it's sort of a 2-in-1... - **Team:** [Hassan Ali](https://cerebralvalley.ai/u/Sem1) - **GitHub:** https://github.com/semswitch-inc/crowdpatch - **Demo video:** https://youtu.be/akwhXcy3MSU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=216 ### 168. Black Box Black Box is a forensic copilot for robots and embodied systems. You hand it a field recording — whatever the platform produces: ROS1/ROS2 bags, raw telemetry logs, multi-camera video, LiDAR/IMU traces, GPS/RTK data, controller logs, or repo context — and it returns a grounded post-mortem: ranked root-cause hypotheses, timestamped multimodal evidence, an NTSB-style PDF, and a scoped code patch. The core insight is that robotics failures are rarely obvious from one log or one operator's account. In our hero case, the operator filed "tunnel caused the anomaly." Black Box's grounding gate cross-checked telemetry, video, and source code and refuted the operator: RTK heading was already degraded 43 minutes before tunnel entry, drive-by-wire never engaged, so the tunnel could not plausibly be the cause. The refutation ships as a ranked hypothesis with its own confidence and patch hint — not as agreement. We built around long-horizon agentic investigation rather than chat. The system includes a FastAPI/HTMX UI with streaming reasoning, a deterministic grounding gate that drops weakly-supported hypotheses, an append-only 4-layer memory stack (case → platform → taxonomy → eval), a frozen 7-class bug taxonomy enforced at parse time by Pydantic, an HITL approve/reject gate before any patch is applied, and Claude Managed Agents for persistent forensic sessions with tool use, checkpoints, time-travel rollback, and operator steering. - **Team:** [Lucas Ercolano](https://cerebralvalley.ai/u/Lucas76hz), [Aayush Shrestha](https://cerebralvalley.ai/u/theaayushstha) - **GitHub:** https://github.com/LucasErcolano/BlackBox - **Demo video:** https://youtu.be/xvMDm-1wKtI?si=AKeD3BcpLPKFglPL - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=219 ### 169. Fethe Network of repo for better codebase and contextual knowledge for agents - **Team:** [Michael Fethe](https://cerebralvalley.ai/u/mfethe) - **GitHub:** https://github.com/mfethe1/memory-network - **Demo video:** https://youtu.be/4f5zY8hbBS0?si=DdUK7cPd1paYUS3Q - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=271 ### 170. ZAALEN Dezimi is a collaborative exchange for datasets — a shared common data environment where researchers, AI teams, and analysts trade cohorts under a credit economy. Three Claude-powered workflows do the heavy lifting: 1. Upload. Claude reads the structure of an uploaded cohort, scores it across 11 quality dimensions (the CALIBER index, 0–10), and pulls live evidence from PubMed, Cochrane and IEEE Xplore. The user signs a sworn declaration (SES eIDAS) and earns credits proportional to the dataset's size and quality. 2. Collect. The user types a natural-language request — "oncology patients over 50". Claude turns it into a structured filter spec via tool use, scans every active dataset in the network, and merges matching rows from multiple cohorts into a brand-new collection. Credits are spent; provenance is preserved. 3. Analyze. One click runs Shapiro-Wilk, Pearson/Spearman correlation, and Ward hierarchical clustering with live SSE streaming and reproducible artefacts. Every action is signed, version-controlled, and auditable. Built in 10 languages on FastAPI + React + PostgreSQL with encrypted at-rest columns, NASA-style code discipline (≤60-line functions, ≥2 assertions per function, bounded loops), and Claude Sonnet 4.6 / Opus 4.7 throughout. - **Team:** [JAVIER DIAZ SANTOS](https://cerebralvalley.ai/u/diazsantos) - **GitHub:** https://github.com/ZAALEN-TEAM/dezimi - **Demo video:** https://youtu.be/JcVd9bNYWmo - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=274 ### 171. evacua Evacua is a voice-first wildfire operations command center for responders. It turns live incident data, responder availability, route advisories, evacuation zones, and environmental risk signals into an auditable command workflow. Operators can monitor active fires on a 3D Mapbox terrain view, dispatch nearby teams, inspect evacuation buffers and route overlays, generate incident briefings, and prepare public alerts with human approval gates. Built with Opus 4.7, Evacua uses AI as an operational planning layer, not just a chatbot. Voice or typed requests produce structured incident action plans, role handoffs for incident/logistics/comms/safety, alert drafts, safety reviews, and ICS-201-style briefs. Each intelligence run is durable and replayable, showing the reasoning trace, before/after operational state, and recommended next actions. The goal is to compress the chaotic first minutes of wildfire response into a calm, responder-focused command surface that helps teams move from scattered signals to coordinated action faster. - **Team:** [Tanush Savadi](https://cerebralvalley.ai/u/tanushsavadi7) - **GitHub:** https://github.com/tanushsavadi/evacua - **Demo video:** https://drive.google.com/drive/foldrs/1N651rhM_a-3cxOAFL_Gc2IVowzvMGUda?usp=drive_link - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=276 ### 172. Samy metref Prior Auth in the Healthcare industry - **Team:** [Metref Samy](https://cerebralvalley.ai/u/pred_aaa) - **GitHub:** https://github.com/predaaaasaaaaaaa/auto-appeal-agent - **Demo video:** https://youtu.be/F--vP9cZ--4?si=PIPYUgSBtInzjW24 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=277 ### 173. Anamnesa Indonesia's national clinical guidelines live in long government PDFs published by the Ministry of Health. A primary-care clinician with a specific question often has one real option: open the PDF on a phone and search through hundreds of pages, sometimes with a patient still in the room. Commercial clinical-decision-support tools solve the interface, but they are English-only and do not cover Indonesian guidelines or the national formulary. I am an emergency physician and GP in Indonesia. Anamnesa is the retrieval layer that workflow is missing, built first for me, and for clinicians who work the way I do. Ask a question in Indonesian. An orchestrator routes it through four agents. A Normalizer (Haiku 4.5) extracts clinical intent from colloquial Indonesian and refuses on out-of-scope or patient-specific questions. A Retriever runs hybrid vector and lexical search against 9,083 chunks from 81 public-domain guidelines, with no language model involved at this stage. A Drafter (Opus 4.7, adaptive thinking, high effort) composes an Indonesian answer with inline citations, can narrow retrieval on its own if the initial chunks are insufficient, and never emits a claim without a source. A Verifier (Opus 4.7, 1M context) independently re-reads every cited chunk and classifies each claim as supported, partial, or unsupported. An unsupported claim gives the Drafter exactly one retry; if the retry also fails, the whole answer is refused rather than shipped. Every citation carries a currency flag, current, aging, or superseded by a newer edition, and clicking one opens the exact page of the source PDF inside the app. - **Team:** [Rifqi Haikal](https://cerebralvalley.ai/u/rifqihaikal) - **GitHub:** https://github.com/0xNoramiya/anamnesa - **Demo video:** https://youtu.be/DJyzwe3ibCI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=4 ### 174. Vultur **Forensic Repo Surgeon** is a portable proof contract for AI-authored frontend repairs. Bring your own agent — our deterministic verifier decides if the bug actually got fixed. Most AI coding tools generate a diff and a confident PR description. None of them produce portable, machine-readable proof that the bug actually got reproduced and fixed in a real browser. Forensic Repo Surgeon ships: 1. A **case contract** — schema describing the bug, target, viewport, repro steps, assertions, and negative control. 2. A **deterministic verifier** — Playwright-driven before/after browser evidence + JSON-configured assertions. 3. A **proof boundary** — `before-verify.json`, `verify.json`, `doctor.json`, `run.json`, `scorecard.json` decide acceptance. Models cannot bypass them. 4. A **live Claude Code agent loop** — Opus 4.7 authors a candidate patch inside the target repository using `Read`, `Glob`, `Grep`, and `Edit` tools through `@anthropic-ai/claude-agent-sdk`. 5. A **side-by-side proof** — fixture-authored patch and model-authored patch are evaluated by the same verifier. The cockpit shows both scorecards, both diffs, the full agent transcript. **Hero result:** On the date-range drawer case, Opus 4.7 read one file, made one edit (`anchor="bottom"` → `anchor="top"`), produced a 13-line diff, and passed the same deterministic verifier as the human-authored fixture (100/100). On the onboarding case, Opus's layout fix scored 92/100 — the verifier honestly caught that mobile touch targets were never widened. The proof boundary did its job. The defensible position is the schema and the verifier, not another agent. Cursor, Devin, Copilot, and Claude Code become consumers of the proof contract instead of competitors to a CLI. Live cockpit: https://forensic-cockpit.vercel.app - **Team:** [Gilang Adam](https://cerebralvalley.ai/u/asynchronope) - **GitHub:** https://github.com/bO-05/forensic-repo-surgeon - **Demo video:** https://youtu.be/LffRhANbTAg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=59 ### 175. Chris Mansfeldt Wildex turns the world into a deck. Photograph any animal — your dog, a pigeon, a snow leopard — and Claude Opus 4.7 identifies the species, generates lore in the animal's first-person voice, and turns it into a collectible playing card. Build decks, battle five themed Trainer challenges or your friends online, with MTG-inspired mechanics (life totals, summoning sickness, type-effective battlefields, signature legendary abilities), and watch your player cards translate into projected square meters of habitat protected via World Land Trust math donating Ad revenue to their "Buy an Acre" program. The goal of this project was to create a positive way to talk, educate and raise money for conservation through the World Land Trust, by using the innovation of AI, in a fun and engaging way. That was achieved on all metrics. Three Anthropic-native moments anchor the experience: Vision capture — Opus 4.7 identifies the species, returns structured JSON with stats / habitat / IUCN tier / lore / conservation note. Cached forever by image SHA-256. Generated species — for any animal outside the curated 33, Opus generates a complete game-ready playing card (clamped to the cost-curve so it can't break balance) and the player can immediately play with it. Unlimited collection size, no manual curation needed. Chat — tap "Talk with this Animal" on any discovered card and Opus 4.7 streams a conversation in the animal's first-person voice, mixing natural-history facts with conservation reality. The Snow Leopard tells you about its blue mountains; the Vaquita explains why there are 10 left in the wild etc. Built solo in 6 days with Claude Code as primary collaborator. React Native + Expo SDK 54 + TypeScript strict + Vercel Edge Function proxy + Upstash rate-limit + budget cap. 33 hand-painted card illustrations + 5 painted load screens + 7 ability VFX. Conservation framing cites WLT "Buy an Acre" math as projected impact at scale when it would be implemented with AD revenue (not donations made — explicit disclaimer in README). Disclaimer: There are several issues that I have solutions for but that are not in the current build. Ofcourse Ad revenue and the automatic fund raising is one of them. The second is photo-detection. At the moment it is possible to take a photo of a photo, and it will register as a card. Although I dont see that as a huge issue because the goal is not "taking photos of the animal" but rather the other side of the loop; learning and raising money for conservation. We would nonetheless build in some failsafes for this like; Claude Geodetecting vs habitat (does it match habitat range of animal), Claude Photo Detection; does it have any elements of being a photo/screen, maybe implementing video mandatory as this is harder to fake etc. - **Team:** [Christian Mansfeldt](https://cerebralvalley.ai/u/TheSwede) - **GitHub:** https://github.com/christianmansfeldt-cloud/wildex - **Demo video:** https://www.youtube.com/watch?v=7uzliLGofoo - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=101 ### 176. Fernando Leyra Have you ever left a meeting thinking: Did I understand that right? Are we about to build the wrong thing? And then carried that uncertainty into the work? That feeling exists everywhere, but in advertising agencies it compounds fast. Strategy, creative, account, and client voices all collide in the same room. Small misalignments turn into expensive revisions. Today, the industry average is five revision rounds per approved idea, up from three in 2007. Communication has not improved. It has become noisier. The hidden costs show up as burnout, team turnover, and a gradual loss of confidence in creative judgment. We fix the problem at its source: the briefing meetings. Our system turns meetings into structured, reliable decision environments for creative teams. By the time the room clears: * The brief is clearly defined and signed off * Client and team conflicts are captured instead of buried * Open questions are visible across strategy, creative, and account * Every decision is documented and traceable The impact is immediate and measurable for agencies: * About 30 hours saved per project * Around $4,500 recovered per project * Roughly $225,000 annually for a 15-person agency But the real shift is human, and this is where agencies feel it most: * The work after the meeting becomes execution, not interpretation * Fewer rewrites between client and creative * Calmer rounds, sharper ideas * Less friction between departments * Less weekend anxiety before deadlines At the core, the model is not the product. It is a commodity. The real value is in the orchestration layer: how prompts are designed, how conversations are structured, and how decisions are captured in a high-stakes creative environment. That is what makes this system trustworthy for agencies, so teams can focus on creating great work instead of debating what the meeting meant. web deploy @ https://clearly.chat/ - **Team:** [Fernando Leyra](https://cerebralvalley.ai/u/leyra) - **GitHub:** https://github.com/fernandoleyra/clearly - **Demo video:** https://www.youtube.com/watch?v=knZUoortIAc - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=155 ### 177. RedWorld AdaalatAI is an AI bench clerk that turns a 200-page Bangladesh court file into a judge-ready bilingual brief and a draft order in five minutes. Bangladesh has 4.65 million pending cases and one judge for every 94,444 citizens. A judge spends four to six hours preparing each case for the bench: reading the FIR, witness depositions, and evidence; cross-checking statements for contradictions; searching 1,484 statutes and the Supreme Court Online Bulletin for relevant precedent; computing whether a procedural deadline like CrPC §167(2) has been violated. AdaalatAI compresses this work to minutes. A custom orchestrator runs nine specialised Claude Opus 4.7 agents , Vision, Classification, Conflict Detection, Precedent Search via MCP, Deadline, Reasoning with adaptive thinking, Translation, Synthesis, and Critic — over a custom MCP corpus of every Bangladesh act and 2,048 pages of SCOB. The output is a bilingual brief in formal Bangla and English plus a signature-ready draft order, with every citation tool-verified by an adversarial Critic agent. The same brief is useful to judges, lawyers, and to plaintiffs who need to understand their own case. - **Team:** [Reduan Khan](https://cerebralvalley.ai/u/Redu) - **GitHub:** https://github.com/redbrofah/adaalat-ai - **Demo video:** https://youtu.be/rET-rgBIbw4?si=J7Z2YdoWfK2epVQ8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=206 ### 178. Photon Consulting Vectimus Warden is the containment layer for AI agents, from Claude Code sessions to autonomous fleets on Amazon Bedrock AgentCore. Three escalation tiers: deny the tool call under 10ms, quarantine the session, terminate the runtime. Works today across eight coding tools and agent frameworks via existing Vectimus adapters. Four model integrations handle the work the deterministic layer can't: Sonnet 4.6 pattern detection across allowed actions, Opus 4.7 fleet investigation, Opus 4.7 policy drafting with sandbox replay, and Opus 4.7 incident narrative. Built solo around full-time work, sixty-plus rescue animals at our sanctuary, and two side businesses. Architecture was thought through on Claude voice mode while mucking out stables, then implemented on Claude Code monitored from the phone between rounds. Four pivots in five days, all in commits and ADRs. Apache 2.0. Built on Vectimus. - **Team:** [Joe Holland](https://cerebralvalley.ai/u/JXavierH) - **GitHub:** https://github.com/vectimus/warden - **Demo video:** https://youtu.be/J9MB53Z3ydo - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=207 ### 179. Vasyl Stanislavchuk Personalized therapeutic stories, written when they matter: - Parents tell about their kid in a short form, then a few short emails - and bring their taste at concept, draft, and style; - Agents write, explore, and render between the gates; - Supervisors keep what ships safe for kids and train the craft corpus so each book starts smarter Pipeline access: https://shelestni.com/judges Login: judges@shelestni.com Password: keep-thinking - **Team:** [Vasyl Stanislavchuk](https://cerebralvalley.ai/u/vasyl) - **GitHub:** https://github.com/vasyl-stanislavchuk/shelestni-app - **Demo video:** https://drive.google.com/file/d/13y8QAi9YYHu0OSdUKq_Cd3Fql55WLCm-/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=211 ### 180. Raghvendra Pratap Singh CampusFlow is the operational coordination layer that academic faculties currently lack. It replaces the Viber-as-a-database, email-thread, Google-Form chaos that medical students live in with a structured, scope-targeted, audited source of truth, without replacing Moodle, e-Index, or the faculty website. The product is manual-first. The default workflow is a typed form. Five record kinds (announcement, schedule change, exam event with optional student confirmation, practical, material pointer) write structured side-table rows plus a shadow announcement so the existing feed and notification fan-out keep working uniformly. A Repeat-for-next-term button on any published record pre-fills the form for the next exam window so professors do not retype the same content every semester. AI is an optional accelerator. A small "Lift from notice" button opens a dialog where staff paste a Viber message or drop a photo of a paper notice. Claude Opus 4.7 with six Zod-typed tools (lookup_course, lookup_group, lookup_location, build_audience_scope, estimate_audience_size, propose_announcement_draft) drafts a structured proposal that lands in the same manual form for human review. No record is ever published without an explicit approval click. What ships: 30-table multi-tenant data model, scope-based RBAC with 19 roles and 22 actions and a deny-by-default can() helper, append-only audit log enforced at the DB level by a Postgres trigger that blocks UPDATE and DELETE, three-layer notification stack with per-importance fan-out, bilingual JSONB on every record, source-of-truth labels with version history, location alias index that resolves "A0" / "AMF A0" / "new building amphitheater" / "amfiteatar u aneksu" to one row, GDPR + ZZPL privacy notice with all eight data-subject-rights surfaces and a subprocessor table. Built solo in one day during a medical-school exam period. - **Team:** [Raghvendra Pratap Singh](https://cerebralvalley.ai/u/rps321) - **GitHub:** https://github.com/rps321321/campus-flow - **Demo video:** https://youtu.be/45I96k1io1o - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=272 ### 181. falsify Teams ship AI accuracy claims that evaporate under scrutiny not from malice, but because the claim was never structured to be falsifiable. A "94% accuracy" number becomes 71% in production and nobody can tell whether the model drifted, the metric got relaxed, or the threshold moved after seeing the data. Falsify is a CLI that forces pre registration: you cryptographically seal the metric, threshold, and dataset with SHA-256 before the experiment runs. Post hoc edits break the hash. CI exits 3. The audit trail writes itself. Deterministic exit codes are the API — 0 pass / 10 fail / 3 tampered. CI gates on the code, humans read the trail. The claim either survives contact with the data, or it didn't happen. Built entirely with Claude Opus 4.7: 5 skills, 2 subagents, 3 slash commands, 1 MCP server. 514 tests passing. - **Team:** [Cuneyt Ozturk](https://cerebralvalley.ai/u/cuneytozturk) - **GitHub:** https://github.com/sk8ordie84/falsify - **Demo video:** https://youtu.be/vVZTNeak5PA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=5 ### 182. Szymon P. Peplinski × Opus 4.7 Feedback enters the product process as evidence. It leaves as opinion. Auditable Design keeps the reasoning intact. The system reads a public Duolingo review corpus, clusters recurring friction, audits it through six canonical design lenses (Norman, WCAG, Kahneman, Osterwalder, Cooper, Garrett), grounds findings in the actual product interface, proposes a direction, and re-audits before handoff. The output is a single markdown brief: measured pain, priority reasoning, validated direction, and full provenance (SHA-256) for every input. The brief is implementation-ready: the demo shows the brief and the corrected mockup side by side. The key contribution of Opus 4.7 appears at the verification stage. During real-product inspection, Opus 4.7 surfaced a pricing inconsistency no upstream layer identified: Super costs 500 gems on one screen and 450 on another — same product, same session. This was not present in user feedback. It was not defined in the heuristic baseline. It required comparing pixels across separate screens. Under identical conditions, Sonnet 4.6 confirmed all baseline heuristics but surfaced no additional findings. Opus 4.6 produced adjacent observations. Only Opus 4.7 produced a new, independent observation. The difference is not accuracy. It is the ability to name what is missing. Cost per cluster: around $1. A qualitative study: tens of thousands. Not louder recommendations — better accountability. Built with Claude Opus 4.7 and Claude Code. - **Team:** [Szymon Peplinski](https://cerebralvalley.ai/u/simm) - **GitHub:** https://github.com/speplinski/hackathon-opus-47 - **Demo video:** https://youtu.be/i9PORz--Aws - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=60 ### 183. Garry Kuwanto Stakeholder feedback for LLM systems is informal: "the classifier feels off on customer reviews," "the assistant is missing the point lately." Translating that into a regression test takes a day of writing benchmarks, hand-checking discrimination, and prompt-tuning toward a moving target. Hypothesize collapses that loop. Given a hypothesis about a system's failure mode and the system's current prompt, it asks Claude to decompose the hypothesis into probing dimensions, generates targeted candidate inputs, runs both the current system and an alternative, and keeps only the inputs where they meaningfully diverge. The output is a small YAML benchmark designed to discriminate between the two — every test exists for a stated reason. The same primitive ships as a CLI, a Claude Code skill that turns natural-language complaints into committed regression tests, and an MCP server exposing the workflow to any compatible host. One emoji-overuse hypothesis: 30 questions tested, 5 discriminating cases, 0 regressions, $0.02. - **Team:** [Garry Kuwanto](https://cerebralvalley.ai/u/gkuwanto) - **GitHub:** https://github.com/gkuwanto/hypothesize - **Demo video:** https://youtube.com/shorts/Skmd0_rhPkQ?feature=share - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=102 ### 184. Sam Holt autolab is an autonomous lab framework with provenance as its foundation. Claude plans and reacts, typed resources execute scientific operations, and every step is written to an append-only hashed ledger. The demo optimizes micromagnetic sensor designs using real physics simulations, comparing Claude as Planner against Optuna TPE on the same workflow, budget, and resource. The key feature is react(): after each result, Claude can continue, branch, replan, retry, ask a human, or stop based on the latest evidence and figures. - **Team:** [Sam Holt](https://cerebralvalley.ai/u/samjrholt) - **GitHub:** https://github.com/samjrholt/autolab - **Demo video:** https://vimeo.com/1186773258?fl=pl&fe=sh - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=156 ### 185. Dare Second Opinion is a multi-agent healthcare AI tool that turns a patient's chart into a cited, fact-checked appointment brief designed for under-researched conditions starting with adenomyosis, a disease that takes 8 years on average to diagnose. A patient drops in their full chart. Five Opus 4.7 agents run a structured consultation: three specialists fire in parallel (OB-GYN, Reproductive Endocrinology, Chronic Pain) consulting from a 32K-token corpus of 2026 guidelines, active clinical trials, and care pathways. A synthesizer composes one cohesive brief. A verifier fact-checks every claim against the corpus and flags anything uncited — green badge = every claim sourced. The brief is research, not advice. Patient charts never persist on disk. The doctor's chair stays the doctor's chair. Output speaks 5 languages. Built solo in 5 days on Opus 4.7's 1M context window with prompt caching. Adenomyosis is the start; the same pipeline answers fibroids, endometriosis, Hashimoto's, long COVID, lupus, POTS — any condition that takes too long, gets dismissed too often, and has a literature too dense for one appointment to hold. The corpus changes; the architecture doesn't. Transparency: "Sandra" is an anonymized real patient profile (with consent). "Maya" and "Helena" are hand-authored composite cases. All five specialists are Opus 4.7 agents not human doctors. Care centers and citations are sourced from public medical literature. - **Team:** [James Olusoga](https://cerebralvalley.ai/u/DareDev256) - **GitHub:** https://github.com/DareDev256/second-opinion - **Demo video:** https://youtu.be/Ha6OVJgm7io - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=157 ### 186. Andrew Diaz Panopticon Live extracts seven biomechanical fatigue signals — recovery latency, serve-toss variance, ritual entropy, crouch-depth degradation, baseline retreat, lateral work rate, split-step latency — from standard 2D broadcast tennis video. No hardware sensors. No depth cameras. No motion capture suits. Just YOLO11m-Pose, a Kalman filter on court-meter geometry, and a 3-pass DAG (forward + RTS smoother + semantic state machine) running on an Apple M4 Pro. The dashboard is a 2K-Sports-style HUD that visualizes the data live — but the FINAL PRODUCT is the downloadable .csv biometric file plus the qualitative match transcripts produced by a three-agent Claude Opus 4.7 swarm (Analytics Specialist → Technical Biomechanics Coach → Tactical Strategist). Tab 3 replays the captured trace with timed pacing, showing the model think, reject hypotheses, and synthesize tactical exploits in real time. Built solo in five days with Claude Code. Twelve project skill packs (cv-pipeline-engineering, physical-kalman-tracking, react-30fps-canvas-architecture, etc.) compose the architecture. MIT license. Live at panopticon-live.vercel.app. The signal is in the infinite pool of pixels on the open internet. Nobody is reading it. We are. - **Team:** [Andrew Diaz](https://cerebralvalley.ai/u/andydiaz122) - **GitHub:** https://github.com/andydiaz122/panopticon-live - **Demo video:** https://www.youtube.com/watch?v=ly2qzRx1mv4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=160 ### 187. Daniel Keck Weather apps make sense out of millions of parameters, but they cannot tell you what the person standing one valley over is actually looking at. The forecast in your pocket is the forecast for everyone in your ten-kilometer cell. VYU turns every phone into a directional weather webcam. A photo plus the compass heading from the device's orientation sensors becomes a view-cone on a shared map, pointing exactly where the camera was aimed. Three Claude Managed Agents on Opus 4.7 do the work, on two clocks. When you submit a photo, a Classifier reads it against a phenomenon taxonomy and a Reconciliation agent fires alongside, checking the image against DWD radar, EUMETSAT MTG infrared and lightning, and the Open-Meteo forecast. What comes out is a verified report. When you ask a question, a Deep Researcher answers it against your own report and a per-cell memory of what other contributors nearby saw. If the answer mentions the cool wind off the lake, it is there because someone there felt it. The agent inherits. One photo earns one question. A fair trade. Tit for tat. - **Team:** [Daniel Keck](https://cerebralvalley.ai/u/dkck) - **GitHub:** https://github.com/of3y/vyou - **Demo video:** https://www.youtube.com/watch?v=1783gSZQ1sQ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=210 ### 188. Patram Like we have Claude Code extensions for IDEs for software developers, my expectation was there should be the same for other knowledge workers mostly working in Docs, Sheets and Ppts. But there isn't. So I set out to build the bare bones skeleton of how that would look, when AI is no more a copilot but actually front and centre in the document editing experience. - **Team:** [Saket Tawde](https://cerebralvalley.ai/u/saketcodes) - **GitHub:** https://github.com/sakettawde/patram3 - **Demo video:** https://youtu.be/V5X0manFUHY - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=275 ### 189. FREDABILA PSON5 is an open-source personalization infrastructure that treats user state as a first-class, structured, auditable artifact — not as free-form notes appended to a prompt. Every profile splits across three strictly-separated layers: observed (what the user said), inferred (what the model deduced, carrying confidence + evidence refs), and simulated (what the engine predicts in a specific scenario). The separation is enforced in the TypeScript type system and in the simulation engine, not just documented. Fifteen MIT-licensed packages under @pson5/* ship an SDK, an Ink/React interactive CLI, a hand-rolled HTTP API with MCP (stdio + HTTP) transports, Neo4j and Postgres storage adapters, a rules-first modeling engine, a simulation engine that emits reasoning + caveats alongside every prediction, and an agent-context projection layer that gives agents consent-scoped, redaction-aware views without ever exposing the raw profile. Two reference apps demo it end-to-end: 1. chat-app — a Claude-powered chatbot that talks to a PSON5-backed profile via 8 agent tools. We shipped a new pson_observe_fact primitive in this hackathon to bridge the gap between open conversation (free-form facts the user volunteers mid-chat) and the structured observed layer, which previously required pre-registered questions for every save. 2. researcher-agent — a Claude Managed Agent that assumes the identity of a fictional Anthropic alignment researcher (Dr. Amelia Kwan, explicitly a composite), seeds a persistent persona profile with 20 structured facts on first contact, and uses pson_simulate to reason about model-building decisions in first person. PSON5's engines run on the host; the agent sees only typed custom tools. The problem PSON5 solves: today's agent memory is opaque, non-portable, and silently collapses evidence with inference. You cannot tell which entries in an agent's memory are things the user actually said vs things the model guessed vs things the simulator predicted, and downstream agents inherit the ambiguity. PSON5 makes that contract auditable by design. - **Team:** [Frederick Abila](https://cerebralvalley.ai/u/fredabila) - **GitHub:** https://github.com/fredabila/pson5 - **Demo video:** https://youtu.be/37kz_rqYIgg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=6 ### 190. Plumbline.ai Plumbline.ai is an AI co-pilot for construction lenders , banks and Construction Risk Management Consultants (CRMCs) verifying monthly draw requests against the plans they're financing. It pairs frontier-grade reasoning (Claude Opus 4.7) with modern field telemetry , drones, IoT sensors , 360° walkthroughs, and fixed-position construction cameras, to turn an inspector's manual workflow into a structured, audit-ready pipeline. Today verification is manual: an inspector drives to site, eyeballs progress, and writes a narrative the loan officer takes on faith. OCC and FDIC have flagged weak Change Order and Schedule-of-Values discipline as a safety-and-soundness issue. An 8-agent Opus 4.7 architecture ingests sealed PDF plans across four disciplines (architectural, structural, electrical, plumbing), the loan's finance plan, the AIA G703 Schedule of Values, and authenticated site evidence. A Managed-Agents Supervisor then investigates findings autonomously and produces a Gap Report with a draw verdict — APPROVE / APPROVE_WITH_CONDITIONS / HOLD / REJECT, citing G703 line items and flagging unapproved scope. Two industries that resisted automation, bank credit risk and field construction , converge here. Agents do the cognition; drones and IoT bring the field witness. Built at the seam between regtech for financial services (my work) and 50 years of practice (my father's). MIT licensed - **Team:** [Asaf Erez](https://cerebralvalley.ai/u/Wutang) - **GitHub:** https://github.com/wutangasaf/hackathonopus - **Demo video:** https://www.youtube.com/watch?v=6i5ijwjFp7Y - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=61 ### 191. Team Credence Every day, developers trust Claude with decisions that matter. They share what they know — and what they don't. "The rate limit is probably around 50, I haven't confirmed it yet." A small admission. An honest one. Then the session grows. Context gets compressed. And silently, without warning, RATE_LIMIT = 50 ships to production. The model just forgot you weren't sure. We named this Epistemic Qualifier Loss (EQL), measured it across 50 compression scenarios, and built Credence to prevent it. Credence is a five-layer epistemic enforcement system built natively into Claude Code. A sub-millisecond faithfulness probe blocks compression when uncertainty is present. A Truth Buffer re-injects unverified constraints before every generation. A Consistency Enforcer fires when a query overlaps a registered uncertain value. A Generation-Time Scanner annotates unverified literals before they ship. A Rust-native PreToolUse gate blocks irreversible tool calls — 98× faster than Python. The Ghost Detector — powered by Opus 4.7 — catches implicit uncertainty with no hedging markers at all. Something no rule-based system can do. As AI agents make higher-stakes decisions, uncertainty is not a weakness to hide — it is the most honest signal a system can carry. Credence is the first enforcement layer built around that principle. 22-tool MCP server. 178 tests. Built entirely with Claude Code. - **Team:** [LAKSHMI CHAKRADHAR VIJAYARAO](https://cerebralvalley.ai/u/Chakrivijayarao) - **GitHub:** https://github.com/Lakshmi-Chakradhar-Vijayarao/credence-ai - **Demo video:** https://www.youtube.com/watch?v=zKEf2k2bIsU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=103 ### 192. PrecisionXxx Convoy is an AI deployment agent that takes a codebase from commit to production with a human-in-the-loop approval model. It scans a repo, identifies the right deployment target, rehearses the build locally, surfaces failures with precise diagnostics, opens and merges PRs for fixes, manages deploy gates, stages required secrets, runs canary releases, and promotes to production only after explicit approval. The problem Convoy solves is the gap between “the code is ready” and “the product is safely live.” For most teams, shipping still involves manual coordination, repeated context switching, hidden environment issues, failed deploy retries, and fragile handoffs between coding, DevOps, and release approval. Convoy turns that messy, error-prone workflow into a guided, auditable pipeline where repetitive execution is automated and judgment stays with the engineer. In our demo, Convoy deployed a real Next.js pharmacy management app to Vercel, recovered from two real failures, resumed from the exact failed stage instead of restarting, validated the fix with live checks, and completed the release with production observation and health verification. - **Team:** [Edward Twumasi](https://cerebralvalley.ai/u/PrecisionXxx) - **GitHub:** https://github.com/teckedd-code2save/convoy - **Demo video:** https://youtu.be/5btzce8adeE - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=158 ### 193. WestCoast Automation Solutions My project is a client facing onboarding system and dashboard that my clients will use to connect their different accounts like social media, crm, google analytics, and even their own webiste to allow the automation systems i have built for my own family business to work in their business as well. - **Team:** [Sam Alarcon](https://cerebralvalley.ai/u/suaveshot) - **GitHub:** https://github.com/suaveshot/wcas-client-dashboard - **Demo video:** https://drive.google.com/file/d/1dRsVAj3X8ephVI95jRmAkvoXCHe5HlEJ/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=161 ### 194. WebLens User-Generative Experiences What if you didn't have to wait for a web application to build the features you want? Today, users are entirely dependent on digital product owners to shape their experience. Want to filter LinkedIn posts by topic? Compare hotels side-by-side on Booking.com? Merge identical announcements about a new Claude release on social media into one? You either wait or you go without. I asked this question: what if users could build their own web experiences without writing a single line of code? The proposed approach I focused on web applications. This insight was simple but powerful: every web page can be decomposed into minimal, meaningful elements: posts, headers, search bars, cards. Once you can identify those elements you can assign behaviors to them. I built a platform where users create an account, define reusable behaviors, and then at runtime drag and drop those behaviors onto any element of any web page. Every behavior is versioned and stored securely in a central database tied to the user's account. In this way, on any system when login into your account, all those behaviors are immediately accessible to you. For this hackathon, the latest Opus 4.7 model is used in conjunction with Claude Code and V0 to develop such a platform. During rune-time behaviors are created using Claude Code API. Given the limited time, the current implementation is tested against the following roles Three Behaviors We Built Filter: Assign this to a group of elements (like a LinkedIn feed) and it dynamically generates categories letting you then to filter and focus on only what's relevant to you. LinkedIn doesn't offer this natively. Translate: Assign this to any element, select your target language and a translation button appears on every subsequent visit. The behavior persists no reconfiguring needed. Summarize: Assign this to any content element and a summary button appears. One click, and you get the core of what's there. Why Now This is only possible today because of generative models like Claude. The intelligence needed to parse arbitrary web elements, generate dynamic UI, and produce meaningful functional outputs during run-time(e.g., translation, summarization, categorization) is now accessible without a team of engineers behind it. Browsers like Comet and platforms like Atlas are pointing in this direction. We think there's still a wide-open space at the user layer and that's exactly where we're building. - **Team:** [hojjat rakhshani](https://cerebralvalley.ai/u/hrakhsha) - **GitHub:** https://github.com/hrakhshani/generative-user-experience-hackathon-anthropic/tree/main - **Demo video:** https://drive.google.com/file/d/1_h0x9lNJvaSwaibWgaq7QDuQxqj-GFgj/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=212 ### 195. Ricardo Garcia Character Chat is an exploratory study into how Claude Opus 4.7 can power a new generation of historical and educational tools. Users hold real-time, in-character conversations with historical figures — Frida Kahlo, Sigmund Freud, Simone de Beauvoir, and Salvador Dalí — each rendered as a pixel-art avatar with reactive emotes and a themed environment that reflects their world. The problem it solves: History and humanities education often feel static — textbooks, lectures, and Wikipedia pages. Students rarely get to engage with the figures they study. Existing chatbots either break character instantly or default to a flat, encyclopedic tone that strips away personality, contradiction, and emotion — the very things that make these figures worth knowing. The approach: Character Chat uses Opus 4.7's adaptive extended thinking to match cognitive effort to the question being asked. Casual greetings get a fast, reflexive reply (thinking disabled). Medium-depth questions trigger a 2,000-token thinking budget. Genuinely philosophical prompts ("what is suffering in art?") unlock a 6,000-token deep-thinking pass, so Frida actually thinks before answering — the way she would have. Combined with structured stage-direction markers parsed by the frontend, the avatar reacts emotionally in sync with the response. It's a study, not a final product: a testbed for how Claude's reasoning capabilities can transform passive historical content into living, in-character dialogue for classrooms, museums, and cultural institutions. - **Team:** [Ricardo Garcia](https://cerebralvalley.ai/u/vellent) - **GitHub:** https://github.com/ricgarcas/character-chat - **Demo video:** https://youtu.be/iHfslqbO37A - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=278 ### 196. Code Diggers NurXai is a Chrome extension and web app that helps users write better replies on X in seconds. Instead of giving generic AI-sounding comments, it generates short, contextual, human-like replies based on the actual post, its tone, and, on higher plans, even attached images. The problem we noticed is simple: most people know replying consistently on X is one of the best ways to grow, network, and stay relevant, but writing good replies at scale is mentally exhausting. Existing AI tools usually feel robotic, repetitive, overly polished, or full of obvious AI tells. NurXai solves that by focusing on naturalness first. It generates multiple reply options that feel casual, platform-native, and usable right away. Users can regenerate for fresh angles, choose different writing styles, and even save project-specific context so replies feel informed rather than generic. this product already includes account-based access, subscription logic, usage tracking, admin controls, context memory, and a working extension-to-web authentication flow. our goal is to make a real creator tool people can use daily on X. Latest info: Uploaded extention on chrome store and it's now in pending review. maybe 12h to 1day it will be fully live. - **Team:** [NURRABBY SHUVO](https://cerebralvalley.ai/u/nurw3b) - **GitHub:** https://github.com/0xnurrabby/nurxai - **Demo video:** https://youtu.be/r2l0pwDzAro?si=chyFKwMelaNODeAC - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=7 ### 197. Cynux Ujian SD turns a single sentence — "Pendidikan Pancasila, kelas 5, sumatif tengah semester, tema sila ke-3" — into a print-ready exam sheet: 20 questions, mixed types (multiple choice, complex MCQ, true/false, short answer, essay), each with an answer key and a teacher-facing pembahasan (explanation). One click → PDF on A4, ready for the photocopier. THE PROBLEM Indonesian SD (elementary) teachers spend 4–8 hours every week hand-writing exam sheets in Microsoft Word. They reuse and recombine the same question banks every term because building from scratch is too slow. The result is uneven quality — items that are too easy, vocabulary above grade level, or explanations written in academic Indonesian no 10-year-old can follow. Generic LLM chatbots produce text that sounds like an exam but fails on the constraints that actually matter: Kurikulum Merdeka phase alignment, learning outcome codes (CP/TP), age-appropriate vocabulary, and proper KD coverage across Bloom's taxonomy levels. WHAT UJIAN SD DOES DIFFERENTLY - Curriculum-grounded generation — the system reads from official Kurikulum Merdeka PDFs (parsed into per-CP/TP units) and ties every question to a real learning outcome, not free-form generation. - Banned vocabulary list for SD level — Opus 4.7 is given an explicit blacklist of academic Indonesian words that shouldn't appear in grades 1–6 material, ensuring age-appropriate language. - Pembahasan that teachers can read aloud — explanations are written for a 10-year-old listener using respectful but simple Indonesian, not dense answer-key prose. - Mixed question types in one sheet — PG (pilihan ganda), PG kompleks, benar-salah, isian singkat, and uraian, properly distributed across difficulty levels following Bloom's taxonomy. - Two-pane preview — the same JSON drives both the on-screen preview and the print PDF, so what teachers see is exactly what gets photocopied. - History + duplicate-as-draft — every generated exam is saved. Teachers can fork last term's sheet, swap the topic, and regenerate without losing the original. - Print-first design — A4 layout, photocopier-safe fonts, proper margins, header with school name/class/semester fields. This is built for Indonesian school reality where 90%+ of exams are still paper-based. TECH STACK Frontend: React 19 + Vite + TanStack Router + Tailwind v4 + Radix primitives Backend: Hono v4 on Node 22 + Effect-TS service layers + Drizzle ORM (Postgres) + better-auth (Google OAuth) AI: Claude Opus 4.7 via official SDK with prompt-cached curriculum context and structured output via Effect Schema validation Monorepo: pnpm + Turborepo with shared Effect Schemas as single source of truth Deploy: VPS + Caddy + Docker, live at ujian-sekolah.faldi.xyz IMPACT A working version of Ujian SD saves an Indonesian teacher 4–8 hours per week on exam creation alone. With 2.9 million SD teachers in Indonesia, even modest adoption addresses a massive, underserved need that no existing edtech product properly solves — because none of them understand what a printed Indonesian exam sheet actually needs to look like. - **Team:** [Naufaldi Rafif](https://cerebralvalley.ai/u/Cynux) - **GitHub:** https://github.com/naufaldi/teacher-exam - **Demo video:** https://youtu.be/5bmG_bbl_MA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=62 ### 198. CivicServant Your personal civic chief of staff. Corporations have lobbyists tracking every bill, meeting, and rule that affects them. Ordinary residents have nothing. Revere is the digital version of Paul Revere's ride: while you sleep, it watches your city council, school board, and state legislature, filters every item through a personal civic fingerprint built from a 10-minute conversation, verifies every factual claim against source, and delivers a 5-minute morning briefing of only what affects your life — with receipts on every claim. The same Austin City Council meeting produces materially different briefings for two real residents three blocks apart. A renter-parent and a small-business owner click on the same rezoning item and see different reasons it matters: housing-cost weight 0.9 versus small-business-permitting weight 0.9. Same record, different lives. When action makes sense, Revere drafts three response variants — direct, measured, persuasive — and runs an adversarial critic loop with three rotating personas (council staffer, opposing constituent, press shop) to harden each draft. You review, edit, send yourself. Revere never autosends. Election Briefing extends the same architecture to upcoming races: candidates cross-referenced against your stated priorities, every promise tagged with a "can they actually do that?" authority badge sourced from the office's charter. Multi-jurisdiction and multi-level by design. Built on Opus 4.7 — high-resolution vision (parcel-level zoning maps), self-verification rigor (every claim source-checked before it reaches you), long-horizon agent orchestration. Harness-agnostic architecture: skill packs portable, sessions logged for traceability, outcome grading via the verification loop. Civic information, personal. Civic action, yours. - **Team:** [Swarit Srivastava](https://cerebralvalley.ai/u/Swarit1) - **GitHub:** https://github.com/swarit-1/revere - **Demo video:** https://www.loom.com/share/544b9b079e98461abfb7c15f0a5fe000 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=279 ### 199. Assistiv Assistiv addresses the UK's Missing Middle, millions of older adults living independently but at risk of sudden decline. Existing care technologies arrive too late, and feel like a threat to identity for people who value their independence. We built an open door: a voice-first, unhurried conversation powered by Claude Opus 4.7, hardwired for unconditional positive regard. Using Motivational Interviewing principles, it helps people reflect on what matters to them. Instead of weeks of navigating fragmented services before anyone has heard them, Assistiv offers a respectful twenty-minute interaction in which the person stays in control, including the ability to say no, which the system treats as a valid and successful outcome. Opus 4.7 powers four parallel tasks for every answer: warm reflection, hostility detection, crisis detection, and adult safeguarding. The system generates a personalised Staying Well Plan grounded in the person's own words, alongside an optional GP Summary for those without advocates. Single-file, privacy-first, no data retained beyond the session. One conversation at a time, with care, Assistiv supports independence earlier. Before crisis. On the person's terms. So that those in the Missing Middle are not locked out, and not alone. - **Team:** [Simon Legrand](https://cerebralvalley.ai/u/simonlegrand) - **GitHub:** https://github.com/silegrand/design-your-life - **Demo video:** https://youtu.be/op3rXkpzbHs - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=8 ### 200. EasyBizzy by Dhrumil Barot Easybizzy is a compliance and incentive intelligence platform for Indian small businesses. From a single sentence about your business - "I run a 22-loom powerloom unit in Kolhapur" - it returns every rule that applies, every government scheme you qualify for, and every gazette amendment that matters. Sorted by deadline. In your language. With every claim traceable to the original document. Indian manufacturers face 26,134 unique compliance obligations (TeamLease RegTech, 2025), spread across 47 central ministries, 36 state gazettes, and 4,000+ municipal bodies - with 9,331 changes in the last year alone. Most can't read them all and can't afford lawyers who can. The result: missed obligations, missed incentives (₹15 lakh ATUFS subsidies, 50% PowerTex grants), and stalled growth. The data already exists; it's just unread. Built on Claude Opus 4.7 with four specialized agents running in parallel — profile extractor, rules matcher, scheme matcher, and a persistent watcher — grounded against a hand-curated YAML corpus of Maharashtra textile-belt rules and central government schemes. The watcher monitors the gazette continuously and surfaces only amendments relevant to the user's structured profile. The translation layer renders summaries, reasoning, and recommended actions in 12+ Indian languages — while keeping citations in the original gazette language, because compliance is sacred. The architecture generalizes across verticals and states. The hackathon scope deliberately stays narrow: small food-processing and textile MSMEs in Maharashtra. The design rule throughout: no vibes. Receipts. Live: easybizzy.vercel.app - **Team:** [Dhrumil Barot](https://cerebralvalley.ai/u/dhrumilbarot) - **GitHub:** https://github.com/barotdhrumil21/easybizzy - **Demo video:** https://youtu.be/eMWVwgnDVl0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=63 ### 201. Abdelmounaim JournalShift is a tool-augmented agentic app: two Opus 4.7 agents running on Anthropic's Managed Agents platform, a Smart Journal Finder that queries OpenAlex, Crossref, and SCImago to rank candidate journals, and a Template Adapter that mechanically rewrites the LaTeX against the user-supplied target template, drive a deterministic Python pipeline that turns a manuscript into a submission-ready package, with Haiku 4.5 handling smaller single-shot tasks (LaTeX preamble repair, reviewer rationale, soft-rule compliance, readiness notes). - **Team:** [Abdelmounaim KERKRI](https://cerebralvalley.ai/u/Profmoun3im) - **GitHub:** https://github.com/abdelkrk/Journalshift - **Demo video:** https://www.youtube.com/watch?v=BVOo4sAhbqo - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=104 ### 202. Orbital Flight School Orbital Flight School is an agentic, gamified textbook for spacecraft dynamics — the subject I teach at the University of London. It tackles the problem of existing textbooks — physical or digital — that are marginally interactive but do not adapt to the student: they cannot tell where the reader is stuck or what should come back tomorrow. Orbital Flight School does more than invert learning; it teaches engineering science in a playful manner. Students learn from this textbook by *playing* — Level 1 wraps a Hohmann-transfer rendezvous in a Mario Kart-like racing game — and by working with agents that personalise their study. Inside the textbook, students can highlight and annotate any passage; these annotations are read by Orbo, an Opus-4.7 teaching assistant who suggests connections, drafts spaced-repetition system (SRS) cards so ideas return before they're forgotten, and provides feedback on the answers. A second managed agent — LISA (LIve SAtellite tracker) — pulls live orbital data from CelesTrak / NASA NTRS / Space-Track and links the student into a 3-D debris visualiser (https://space-debbie.vercel.app), so the textbook also opens out into research: LISA references existing missions and debris while Orbo surfaces ideas and facts on demand. The visualiser is an ongoing research project of mine. I'm deploying this in my next-semester offering. Alongside the project I'm publishing a sibling repo — my hackathon stretch goal — that lets anyone convert any PDF into the same agentic textbook format; they get the SRS + Orbo TA out of the box. I call it Booksmart (https://github.com/angadhn/booksmart), a learning technology that adapts to students instead of students having to adapt to the technology and education system. - **Team:** [Angadh Nanjangud](https://cerebralvalley.ai/u/angadh) - **GitHub:** https://github.com/angadhn/orbital-flight-school - **Demo video:** https://youtu.be/XJh-zc94OtQ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=163 ### 203. Saad EL BABIDI Interior architects in France designing offices for banks, law firms and consulting groups are leaving money on the table. Each one handles only one or two clients a month — not because the demand isn't there, but because they spend 60 % of their time on iterations a software should do for them : drawing in AutoCAD, rendering in SketchUp, checking French codes manually, building mood boards in Photoshop, for every variant, every iteration, every client meeting. So when a fifth qualified prospect calls, ready to sign, they get added to a waiting list. Or lost. Archoff is the quiet co-architect that takes back those weeks. From a single client brief, it produces a sourced functional programme, three 3D test-fit variants pushed live into SketchUp, a curated mood board sourced from real catalogues like Vitra and Herman Miller, a sourced argumentaire backed by neuroscience research and French regulation, an 18-slide magazine-grade client presentation, and a dimensioned DXF dropped straight into AutoCAD for the engineering hand-off. The architect refines through a chat that doesn't just answer — it acts : add a zone, move a wall, regenerate a variant. What used to take six weeks now takes minutes. The architect can finally answer the fifth call. - **Placement:** Finalist - **Team:** [Saad EL BABIDI](https://cerebralvalley.ai/u/Saadzw) - **GitHub:** https://github.com/Saadzwak/design-office - **Demo video:** https://youtu.be/JetiYGxOCT4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=164 ### 204. CTRL_ALT_DESIGN I ship and manage design systems for clients. BELLA is the synthetic stand-in I built for this demo — I can't show real client data, but the workflow is identical. Every system I onboard has context locked in team members' heads. Every morning I check ten tools by hand — Storybook, Figma, Zeroheight, three platform repos, Jira, Slack, Notion — to find where parity drifted overnight. The decisions get made on gut, because the data's too scattered to use. Most "design system tools" are static doc sites. They describe the system; nobody watches it. CHIP is different. It's a control room where an agent watches the system and proposes the diff before the parity drifts. When drift is detected, CHIP files an audit, drafts a roadmap, and waits. I review. I approve. CHIP files the Jira tickets, notifies the platform leads, and writes the diff to the audit ribbon. Every action traceable. No silent merges. Behind the Ask CHIP input is a Claude Managed Agent on Opus 4.7 with the Notion MCP wired in. When I ask "what does Vitaly say about dashboards", it queries my actual research library — 47 sources from six primary voices on AI-ready design systems. The dashboard runs the loop on itself. The loop closed. - **Team:** [Elleta McDaniel](https://cerebralvalley.ai/u/CTRL_ALT_DESIGN) - **GitHub:** https://github.com/emcdanie/chip - **Demo video:** https://www.loom.com/share/ab0ac2410e3947b49efb5b8166dcd63a - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=165 ### 205. Ufuk Karaca Spine is an editorial workspace that turns a novel into a navigable graph. Scenes are commits, plotlines are branches, characters are contributors with line-level blame. You drop in a manuscript and get back its structure - which threads weave together, which fizzle out, which characters drive which arcs. **The problem.** Developmental editors read hundred-thousand-word drafts to find structural issues: subplots that get dropped, character arcs that flatten, pacing that sags between climaxes. It's a job that lives in the editor's head - they read for weeks, take longhand notes, and slowly assemble a structural picture. The knowledge is hard to share, hard to teach, and impossible to query later. **What Spine does.** It uses Opus 4.7's 1M-token context window to hold the entire manuscript - and treats Claude as a literary editor, not a chatbot. The hero feature is the **agentic editorial brief**: click one button, Opus 4.7 reads all 321 scenes of *Pride and Prejudice* through 1M context, calls eight read-only tools to verify each claim against the literal scene text, and writes a real editorial brief - the kind a senior dev editor would hand back. *"A structurally exemplary three-act courtship novel. Volume one plants every bomb. Volume two detonates them at Hunsford. The looseness is in the post-crisis stretch where the Lydia rescue is processed in many small repetitive beats."* That's a real sample from a real run. Specific. Cited. Verifiable. **Other surfaces.** A tension ribbon along the top of the workspace shows every scene's importance over time, color-coded by plotline - the shape of the book at a glance. An "Ask the manuscript" surface uses the same agent loop with UI-control tools (`highlight_character`, `focus_plotline`, `open_scene`) so the model literally drives the workspace as it answers. A plot graph view surfaces orphan threads (never connect) and unresolved threads (trail off without merging). Characters can be talked to in their own voice, with full-book context. **Who benefits.** Developmental editors at agencies and publishing houses. Authors revising long-form drafts. MFA students. Literary scholars who want to query narrative structure quantitatively. The thing that took weeks now takes seconds - and the structural read is shareable, queryable, and persistent. **Stack.** Next.js 15, React 19, TypeScript strict, SQLite via better-sqlite3 with raw SQL. Three Claude tiers: Haiku 4.5 with prompt caching on the per-scene ingest hot path, Sonnet 4.6 for mid-weight reasoning, Opus 4.7 with adaptive thinking on the synthesis surfaces. Pioneer (Fastino's GLiNER2) for character disambiguation NER. Tavily for real-world research when Ask triages a question to it. Built end-to-end with Claude Code. - **Team:** [Ufuk Karaca](https://cerebralvalley.ai/u/ufukkaraca) - **GitHub:** https://github.com/ufukkaraca/spine - **Demo video:** https://www.loom.com/share/a8b92a4a02634110b5eb6c94b68cdeca - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=214 ### 206. Bonsai Bonsai is an agent that manages your personal expenses. People leak money across insurance, rent, credit lines, taxes, medical, utilities, and subscriptions, but the savings on any single bill aren’t worth the time to investigate. So nobody does. Bonsai automates the work end to end. Upload a bill, and the agents audit it, drafts an appeal, and negotiates by email or phone. Four agents split the job: 1) The analyzer finds opportunities to lower your bill. 2) The contact resolver finds the billing department’s email and phone. 3) The negotiation agent emails or calls the provider and handles the back-and-forth. 4) The comparison agent runs in the background, hunting for cheaper providers you could switch to. Try it: https://bonsai.firebaystudios.com - **Team:** [Garrett Cahill](https://cerebralvalley.ai/u/Garrettc) - **GitHub:** https://github.com/garrettc23/bonsai - **Demo video:** https://youtu.be/JlCOw8Q5-Ds - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=280 ### 207. VOID TECHNOLOGY Project Description Misologist is an AI fermentation assistant for home miso makers, small breweries, and fermentation researchers. It helps users diagnose risky batches from photos, monitor fermentation over time, and translate traditional craft intuition into clear scientific explanations. Using Claude for vision, reasoning, and long-context synthesis, Misologist can explain whether a batch is safe, what likely went wrong, what action to take next, and how conditions across time affect the final result. It also helps users design recipes from target flavor goals, turning “I want a sweeter, less acidic miso” into practical fermentation parameters. Problem It Solves Miso fermentation is slow, high-stakes, and hard to debug. A single batch can take months, and if something goes wrong, the loss is often irreversible. At the same time, much of the best fermentation knowledge still lives in tacit craft wisdom rather than structured, accessible guidance. Misologist solves both problems: it reduces batch failure by giving timely, explainable diagnosis and next-step recommendations, and it preserves expert know-how by converting traditional rules of thumb into explicit fermentation science that more people can use. - **Team:** [Takuro Tsujikawa](https://cerebralvalley.ai/u/takurot) - **GitHub:** https://github.com/takurot/misologist - **Demo video:** https://www.youtube.com/shorts/qFmees4hxd8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=9 ### 208. K&S Venture Group neo-triage is a real-time, hybrid-classifier triage system for Near-Earth Object follow-up. When the Vera C. Rubin Observatory enters full operations, the planetary-defense community will be hit with ~130 NEO candidates per night — about 8× the current Minor Planet Center load. Roughly 8% are genuine NEOs; the rest are main-belt asteroids, satellites, and tracklet artifacts. Human observers cannot read every tracklet AND still take the right shots tonight. The architecture is hybrid by design. A calibrated Bayesian ranker (sklearn GBM) scores every fresh MPC NEOCP tracklet in milliseconds. Claude Opus 4.7 then acts as expert reviewer on top of that — reading the same features but adding context the feature vector cannot carry: observatory bias, blended-source risk, rate-magnitude geometry inconsistency. Opus emits structured verdicts via tool_use (CONCUR / PARTIAL_CONCUR / DISSENT) plus a streamed 150-250 word veteran-astronomer briefing in the operator's UI, with extended thinking visible. Five distinct Opus 4.7 touchpoints: operator briefing, hybrid expert classifier, 2024 YR4 historical replay (re-assessing milestones h+0 through h+168), live-written global alert (cache-bypass for genuine threats), and a Managed Agent loop that broadcasts new tracklets every 2 minutes with reasoning attached. The system cross-validates against JPL Sentry-II, ESA NEOCC Aegis v5, and JPL CAD. The Imminent Impactors Library embeds six historically verified pre-impact predictions (2008 TC3 Sudan, 2014 AA Atlantic, 2022 EB5 Norway, 2023 CX1 France, 2024 BX1 Berlin, 2024 YR4 cleared by JWST 2025) — every coordinate verified against ≥2 published sources. The 2024 YR4 corridor in the app reproduces the actual ESA NEOCC published February 2025 trajectory: Pacific → South America → Atlantic → Africa → Indian Ocean → Bangladesh, ~110M people in the original risk zone. Honest scope: this is a triage layer that cross-validates against production planetary defense systems. It does not replace JPL Sentry-II or ESA Aegis. Phase 2 of the public roadmap integrates real Find_Orb b-plane Monte Carlo. Open source. MIT. - **Team:** [Kacper Saks](https://cerebralvalley.ai/u/Ricko12v) - **GitHub:** https://github.com/Ricko12vPL/neo-triage-hack - **Demo video:** https://youtu.be/0ZP-ID-be8s - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=64 ### 209. RxGuard According to the WHO, medication errors contribute to over 3 million deaths every year, most of them in low and middle-income countries where no pharmacist exists to catch a dangerous prescription. RxGuard is built for those patients. My wife is a pharmacist in Belgium. When I asked her what she does before giving a patient their medicine, she described checking every drug interaction, every contraindication, every patient-specific risk. Most patients in Bangladesh (my home country) never get any of that. They hand over the prescription, get the drugs, and go home. That gap is what RxGuard is built to close. RxGuard is a prescription safety checker for patients in Bangladesh who do not have access to pharmacists due to the critical shortage of qualified pharmacy professionals in the country. A patient photographs a handwritten prescription, answers a few questions about themselves, and RxGuard flags any known dangers and then tells them to return to their doctor. Claude Opus 4.7 does two things: 1. Reads handwritten Bengali and English prescriptions using vision OCR, and 2. Reasons over real FDA and NLM clinical data to detect drug interactions and patient-specific risks. Every flag requires a verified citation from OpenFDA, the NLM Drug Interaction API, or the WHO Essential Medicines List. RxGuard never recommends alternatives, never changes doses, and never stores patient data. It only flags potential risks and recommends returning to the doctor. Our goal is to scale RxGuard beyond Bangladesh, to every country where pharmacists are scarce and patients are at risk. According to the WHO, 2 billion people worldwide lack regular access to essential medicines. We cannot train thousands of pharmacists overnight, but we can put a prescription safety check in every pocket. Starting with Bangladesh, expanding to Africa, South Asia, and every underserved region where a wrong medicine can cost a life. Our goal is clear and simple: no prescription should go unchecked. - **Team:** [Mishkat Chowdhury](https://cerebralvalley.ai/u/mishkat96) - **GitHub:** https://github.com/Mishkat96/Rxguard - **Demo video:** https://youtu.be/K3pdMg2mnaU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=66 ### 210. Hapoel # Ochre — An LLM-First Knowledge System **The problem.** LLMs hallucinate because they read codebases without a map. Documentation rots faster than humans can audit it. Existing wikis are written for humans; agents drown in stale, contradicting prose and cite whatever they find first. **What we built.** Ochre is a documentation system designed to be *operated* by an LLM, not just read by one. The `docs/` tree is the product. Every doc declares an authority class — canonical, supporting, generated, or archived — so an agent knows what to cite and what to ignore. A live registry, an append-only operational log, and bidirectional contradiction edges keep truth visible at retrieval time, not buried in commit history. **The kernel that enforces it.** Four Go helpers validate every change. `docsval` hard-fails CI on schema drift, link rot, missed revalidations, even a stale validator binary. `kernelmath` runs an orthogonal-greedy retriever over Ollama embeddings, with per-doc health scores and Mahalanobis anomaly signals flagging suspect docs before an agent cites them. A learned ridge-fit policy and logistic stop-decider tune retrieval against a held-out 25-query NDCG@5 set, so quality is measured, not asserted. - **Team:** [Tal Elazar](https://cerebralvalley.ai/u/Hapoel) - **Demo video:** https://youtu.be/FYn4F_75ACk - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=105 ### 211. Mike's team Most AI resources are either too shallow (hype) or too dense to understand, leaving a massive "intuition gap" for newcomers. People use AI daily but don't understand the "why," which limits their ability to build or innovate. BRIDGE is an interactive learning ecosystem designed to turn curious users into informed builders. It focuses on three pillars: 1. Theory Mode: Complex concepts explained through high-level analogies and zero jargon. Including a "How an Agent Works" demonstration. 2. Interactive Practice: A sandbox where users build and deploy a "mini-AI" in under 2 minutes. 3. Visualization: Interactive maps of word embeddings to make abstract high-dimensional math tangible. With actual mockups to escalate the project to using agents helping learning features. BRIDGE democratizes technical literacy, providing the foundational "bridge" needed for the next wave of AI engineers and enthusiasts. - **Team:** [Miguel Angel Lopez Munoz](https://cerebralvalley.ai/u/mike4698) - **GitHub:** https://github.com/miguelangellomuresp-dev/OpusHackatonMike - **Demo video:** https://youtu.be/4yHFQznKrlc - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=166 ### 212. ARIA In oil & gas, manufacturing, retail, and most traditional industries, the most valuable operational knowledge isn't in any database — it lives in people's heads. The engineer who remembers why a pump failed in 2019. The technician who modified a valve and never updated the CMMS. You can't reach that knowledge with an MCP server, and you can't RAG your way to it. You have to ask, then wait — sometimes for days. ARIA is a human-in-the-loop agent platform that treats people as a first-class data source alongside tools and skills. When the agent hits the boundary of what any system can tell it, it reaches out directly — over email, phone, SMS, Slack, or Zoom transcripts — to the right human, waits however long it takes, integrates the response, and continues. Every inbound message routes back to the same managed agent session via webhook, so the agent picks up exactly where it left off, days later if needed. The first deployment targets oil & gas FMEA/RCM analysis. ARIA fans out three parallel specialist sub-agents (rotating equipment, static, instrumentation), identifies missing context, batches questions to the analyst pool, escalates up the hierarchy if no one replies, falls back to a phone call when needed, and produces a 4-sheet regulatory-grade Excel report. A process that takes a reliability team three weeks now runs overnight. The same architecture generalises to any industry where institutional knowledge is distributed across people rather than systems. - **Team:** [Areeb pasha](https://cerebralvalley.ai/u/areebp) - **GitHub:** https://github.com/areeb1501/ARIA - **Demo video:** https://www.youtube.com/watch?v=7M7cCDToC8o - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=217 ### 213. Kover Dream Kover Dream is a mobile app powered by Claude that captures children's most ambitious dreams (e.g., "I want to go to Mars") and transforms them into interactive and achievable experiences using simple recycled materials (cardboard, tape, etc.). The AI ​​generates a magical map of missions (draw, build, chat with a mentor). Each physical creation by the child is validated in real time by the visual analysis of Claude Opus 4.7, which then generates a draft response for a human mentor (Human-in-the-loop). Our mission: to restore equal opportunities through imagination. - **Team:** [Bilel Gouja](https://cerebralvalley.ai/u/BibiG) - **GitHub:** https://github.com/BIlelGouja/KoverDream-Hackathon-Claude/tree/main - **Demo video:** https://www.youtube.com/watch?v=U_u6Pj6IOUo - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=10 ### 214. John Knopf Atelier is an AI art director for working photographers. Upload your portfolio, build an Artist Knowledge Base through a short structured interview, and get a Career Dossier with ranked grant / residency / competition / gallery opportunities and submission-ready application materials. The system runs as a long synchronous pipeline of six specialist agents on Claude Opus 4.7. Two of them run on Anthropic's Managed Agents beta: Opportunity Scout — searches twenty-plus curated photography archetypes (landscape competitions, photo-book prizes, photography-specific foundation grants, photo residencies, museum-acquisition tracks, etc.) using web_search + web_fetch, applies eligibility filters, locates past recipients with portfolio image URLs, and persists each opportunity via a custom tool round-trip. Rubric Matcher — the novel core. For each opportunity, it pulls past recipients' portfolio images, normalizes them through Sharp, uploads to the Anthropic Files API, and sends them as image content blocks inside per-opportunity user.message events alongside the photographer's own work in the same Managed Agents session. The agent reads both cohorts visually and scores aesthetic fit with reasoning that cites specific portfolio images and names individual past recipients (e.g., "Hougaard Malan and Max Rive define this prize's pastel-pano formula; your Tahoe Bonsai Rock frame at 2:1 hits exactly that vantage"). The other four agents are direct SDK calls: Style Analyst (Opus vision chunked over the full portfolio, produces a structured aesthetic fingerprint — composition, palette, formal lineage, museum-acquisition signals), Knowledge Extractor (gap-driven structured interview that walks a priority-tiered AKB field list and only asks for what's missing), Package Drafter (institutional-voice writing — statement, proposal, cover letter — drafted in the voice each specific program expects, with a fact-grounding linter that rejects hallucinated venues / fabricated partnerships / made-up dates before they ship), and Orchestrator (composite scoring, ranking narrative, master CV, "why not" blurbs for filtered-out opportunities). Twenty-one curated skill files codify the photography-domain knowledge — controlled aesthetic vocabulary with twelve named precedents (Adams, Sugimoto, Eggleston, the Bechers, Misrach, Mann, etc.), juror-reading heuristics for institutional cohort analysis, opportunity-source registry, voice rules per material type, fact-grounding patterns. Loaded into the appropriate agent system prompts at runtime. The artifact is a Career Dossier — printable PDF and web view — that names what to apply to, what to skip, and why. Each included opportunity carries a fit rationale citing specific portfolio images, plus pre-drafted statement / proposal / cover letter / work-sample selection (12 images with per-image rationale) in the institutional voice that program expects. Each filtered-out opportunity carries a one-sentence "Why not [program]" — saying no with reasons is the product, not a by-product. The dossier surfaces the harsh-truth output prominently rather than hiding it. Architecture decisions worth flagging: the entire app deploys to Vercel — no Cloud Run, no separate worker tier, no long-lived process. Long-running agentic work survives Vercel's 60-second function timeout via a poll-pull-on-read pattern (state in Turso, browser polls a thin /events route, that route reads new events from Anthropic and persists them on every poll, server-side cron handles browser-tab-closed continuity). Image content blocks at session scale bypass the read-tool degradation that kicks in above ~95 mounted files. Sequential per-opportunity dispatch keeps each Rubric turn under context limits. All custom-tool persistence is idempotent against unique partial indices on event_id. Version one is photography-only and single-tenant. The hosted multi-tenant version (email/password accounts, no bring-your-own database or API key) is roughly a day-and-a-half of work after the hackathon — the auth and persistence seams are pre-wired today. Free for working photographers when it ships. - **Team:** [John Knopf](https://cerebralvalley.ai/u/JohnKnopf) - **GitHub:** https://github.com/johnkf5-ops/Atelier - **Demo video:** https://youtu.be/bNnt5r1GBmM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=65 ### 215. Lemonode You describe what you want covered in one sentence: "Chicago weekends with the kids, every Friday at 7am" or "Iran–Israel daily briefing, primary sources, deep" or "Musicals & Live concerts near Paris, every week". A Beat Designer agent (Sonnet 4.6) turns the brief into a structured spec, asking at most one or two clarifying questions only when a dimension is genuinely ambiguous. A Sources Scout agent (Sonnet 4.6) discovers and verifies 10–25 sources spanning at least six distinct domains, leaning on a recipe catalogue (Google News RSS, Reddit JSON, HN Algolia, RSS feedparser) before resorting to generic web search. On the cadence you set, an Editor agent (Opus 4.7) reads the latest content, applies your beat's learned relevance rules, and drafts a polished issue that ships to your inbox via Mailgun. The whole editorial loop runs inside Claude Managed Agents - three sub-agents, two memory stores per beat plus one global, an environment with python3.12 + node20 + feedparser + beautifulsoup4 + requests. Wasp owns auth, the dashboard, scheduling (pg-boss), email dispatch, and a live transcript view that polls agent.events so you can watch the Designer think. The split is deliberate: Wasp's Postgres holds product data; memory stores hold the agent's brain. Three example beats - a hyper-local one, a topical-global one, and a crossed beat - prove the personalisation is real and not a wrapper. Live Demo: https://pb.lemonode.pl Sample Account: login: hello@lemonode.pl pass: claude123!@# You can register with your own email & configure your own personal beats, or can login on sample account with already configured beats, to preview everything - from scheduling, adding new beats, to exploring already delivered emails. - **Team:** [Pawel Szpiczakowski](https://cerebralvalley.ai/u/panpawka) - **GitHub:** https://github.com/panpawka/personal-beats - **Demo video:** https://youtu.be/IO-IemayFzg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=106 ### 216. Abdullah Habbar Isghaa (إصغاء — Arabic for “to listen attentively”) is an AI-powered telehealth platform for Arabic and English mental health sessions. It transforms the clinical workflow for therapists and patients across three integrated apps: Patient Portal — Patients log in via OTP, book appointments through a conversational AI assistant (Claude Sonnet), join video sessions, and receive post-session care. The booking flow features natural-language date/time selection with interactive UI chips and booking confirmation cards. Provider Portal — During live video sessions, providers get a real-time clinical copilot (DSM-5 signal detection, risk assessment, suggested follow-up questions), live Arabic/English transcription tied to the HMS mute state, and the flagship AI Clinical Scribe — Claude Opus 4.7 with extended thinking that generates a complete SOAP clinical note during the session, ready to sign and export as a PDF the moment the call ends. Admin Portal — Includes an anonymous provider quality assessment system using the validated WAI-SR (Working Alliance Inventory) framework + 4 clinical safety items, auto-generated after every session using Claude with strict patient anonymization rules. The platform integrates with ERPNext (patient records), 100ms HMS (video), Deepgram (English transcription), and OpenAI gpt-4o-transcribe (Arabic transcription), all proxied through a secure Node.js server that prevents client-side key exposure. - **Team:** [Abdullah Habbar](https://cerebralvalley.ai/u/habbar) - **GitHub:** https://github.com/ahabbar/isghaa-intake-assistant - **Demo video:** https://youtu.be/dQuG4S8EPBw - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=167 ### 217. Mosquit MosQuit is an AI-powered command center where four competing agents debate in real-time to decide which neighborhoods to fumigate during dengue outbreaks. When cities face limited resources - 4 teams, $30k budget, 48-hour window - MosQuit's agents argue from different perspectives: the Epidemiologist pushes for aggressive intervention, the Budget Officer questions costs, Operations ensures logistics work, and Public Risk considers community impact. Built with Claude Opus 4.7, MosQuit showcases: - Extended thinking for transparent agent reasoning - Multi-agent orchestration with visible debate and disagreement - Prompt caching for efficient 4-agent coordination - Real satellite data integration (Sentinel Hub, Open-Meteo) The result: a consensus plan with documented trade-offs, a 7-day outbreak simulation, and the answer to "how many lives did we save?" - **Team:** [Gabriel Olarte](https://cerebralvalley.ai/u/olartgabo) - **GitHub:** https://github.com/olartgabo/MosQuit - **Demo video:** https://youtu.be/0CNwGZe_TIU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=281 ### 218. Toki ClawBus is a minimal coordination protocol for multi-agent Claude Code — a trust layer that makes delegation, approval, and audit first-class protocol events instead of UI callbacks. It lowers the trust barrier to multi-agent Claude Code by making every task, approval, and result inspectable after the fact. The itch came from running multiple Claude Code sessions across a 5-node cluster for months: every coordination layer I built reached for workflow engines bigger than needed, while hand-offs still happened by copy-pasting context between tabs and approving diffs on vibes. ClawBus replies with five message kinds (task / result / approval- request / approval-decision / log), an append-only store with causal parent links, and an adapter interface four methods long. Three adapters ship — File, SQLite, Discord; the last lets a human ✅ reaction become an approval-decision without the worker knowing which transport it's on. Same protocol, different transport, identical audit trail. Built on the Claude Agent SDK with Opus 4.7, every file mutation routes through an approval-request — a fully inspectable message carrying the proposed diff, severity, and rationale. The gate doesn't pause on vibes. MIT TypeScript, 27 tests on CI, three captured end-to-end runs auditable in five minutes. - **Team:** [Toki](https://cerebralvalley.ai/u/toki) - **GitHub:** https://github.com/tokimwc/clawbus - **Demo video:** https://youtu.be/dS2q9b5zXsU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=11 ### 219. Demonurgo Briefy is a B2B SaaS for content marketing agencies. The core problem it solves: agencies waste too much time coordinating clients, briefs, and content calendars across scattered tools (spreadsheets, emails, Notion, etc). Solutions 1. AI brief generation Claude generates structured briefs from demand context (objective, tone, channel). Inline editing — no page navigation required. 2. AI monthly planning Claude generates a full editorial calendar for a client's month. Users accept, reject, or bulk-convert suggestions into real demands in one click. 3. Per-demand AI chat Multi-turn conversation with full history scoped to each demand. Writers and managers can ask questions, request variations, or refine creative direction without leaving context. 4. Centralized client & demand management Client profiles with monthly post quotas, important dates, and history. Kanban board with drag-and-drop to track each content demand through its workflow. 5. Real-time collaboration Kanban and comments update live via WebSocket (Laravel Reverb). Real-time notifications with badge counter. 6. Team management with access control Email invites, role-based access (owner / admin / collaborator), multi-org support. - **Team:** [Pedro Augusto](https://cerebralvalley.ai/u/demonurgo) - **GitHub:** https://github.com/demonurgo/briefy - **Demo video:** https://youtu.be/V1zHTSnipgc - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=29 ### 220. GetVibing.ai Symphony is an orchestration and governance platform for agentic business processes. Design a workflow in a visual editor — or describe one in plain English and let a Flow Builder agent create it — and publish it as a live service callable from Cowork, Claude.ai, Slack, cron, webhooks, or other flows. Build once, run from any channel. Claude managed agents are the runtime. Every workflow node is a persistent agent with its own instructions, tools, and integration surface. A natural-language concierge resolves "log the Acme deal — $48K" into the right agents, tools, and Salesforce records without the user ever touching the CRM. An App Factory generates React forms from live Salesforce metadata on demand. A Flow Builder meta-agent produces, validates, and publishes complete workflows during its own session — agents building agents. Governance is the architecture, not a feature. Append-only audit logs, token budgets enforced at runtime, fail-closed human gates delivered through Slack, and versioned agent configs so any historical execution can be reconstructed. Symphony is the layer between "we have agents" and "we trust agents with our business." More detail and perspective here: https://docs.google.com/document/d/1GzYdESYQn_v47PXHj779pTAaj2rsBiY39JGmzjOxaSg/edit?usp=sharing - **Team:** [Ramin Hedayatpour](https://cerebralvalley.ai/u/GetVibing) - **GitHub:** https://github.com/RaminAgentic/Symphony - **Demo video:** https://youtu.be/kpGivH9FBMQ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=67 ### 221. Takafumi Mizuguchi At 2 AM a parent's child has a fever. Three options: call a professional, drive to an open pharmacy, or wait until morning. The right one needs medical judgment the parent doesn't have. Japan's pediatric advice line takes 1.24M such calls a year — and its own ministry flags eleven structural gaps. Yorukusu is a SwiftUI iOS app that loads a pharmacist-authored Agent Skill (red-flag screening, OTC-only logic, parent-facing safety rationale), reconciles against a device-local Personal Health Context, and asks Claude Opus 4.7 to return one of three answers: CALL the right emergency number, GO to a pharmacy with a pharmacist on duty, or WAIT with explicit warning signs. Yorukusu does not diagnose — pharmacist-informed, not pharmacist-replacing. - **Team:** [Takafumi Mizuguchi](https://cerebralvalley.ai/u/guttyo) - **GitHub:** https://github.com/Guttyo/yorukusu - **Demo video:** https://youtu.be/FdMz9ZeP1Dg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=107 ### 222. Youssef Ahmed Afify MedMamba is an autonomous medical-image diagnostic system: drop in any image and the model decides for itself which of four imaging modalities it is (histopathology, dermoscopy, blood-cell microscopy, or retinal OCT), runs the matching diagnostic head, generates an SSM-GradCAM saliency map, and asks Claude Opus 4.7 to produce a clinician-style explanation of what it's looking at — with no human ever specifying the modality. The problem: standard multi-task medical classifiers fail on an unspoken assumption — a human must tell the model what kind of image it is at inference. That makes them useless for any real triage scenario where the modality is the question itself. The architectural contribution: a Supervised-Contrastive auxiliary objective trained alongside the multi-task heads forces the VMamba backbone to learn modality-separating geometry. After training, per-modality prototypes are stored inside the model. At inference, an unknown image self-routes via 1-NN cosine similarity in feature space, then the matching head fires. Two objectives (fine-grained class CE + coarse-grained domain SupCon) operate at orthogonal scales, so satisfying both costs essentially no per-task accuracy. Trained results across 4 MedMNIST+ datasets: avg F1 = 0.90 (PathMNIST 0.99, BloodMNIST 0.99, OCTMNIST 0.92, DermaMNIST 0.70 — class-imbalance bound), 93/93 tests passing. The Opus 4.7 contribution: the autonomous prediction (modality + class + confidence + top-k) and the GradCAM heatmap are sent to Opus 4.7 alongside the original image. Opus 4.7 returns a structured JSON explanation — focus_region, finding, differential, confidence_note, follow_up — using the right vocabulary per modality (histopathology, dermoscopic, cytomorphological, or retinal-OCT terms). Because Opus 4.7 sees both the image AND the heatmap, it surfaces a known failure mode pure-softmax can't: high model confidence + diffuse heatmap = flag for manual review. Provenance: every line was written collaboratively with Claude Code. The SupCon contrastive feature itself was planned, implemented, tested, and verified by a single Opus 4.7 Claude Code agent session — 7 files changed, 17 new tests added, 93 tests green on first run. - **Team:** [Yusuf Afify](https://cerebralvalley.ai/u/Afify) - **GitHub:** https://github.com/yusufafify/interpretable-med-mamba - **Demo video:** https://drive.google.com/file/d/1-QHEUw5E0SZGlT99_jXfTVJx8qBEoScQ/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=168 ### 223. Nana Chon nanaOS is an AI-native design system where code and contracts are the single source of truth. Instead of splitting design intent across Figma, code, documentation, and Storybook, nanaOS turns tokens, semantics, accessibility rules, component states, and system architecture into explicit contracts that both humans and AI agents can read and evolve. The project uses a 5-layer architecture: primitives, semantics, rules, components, and patterns. Each component ships with SCSS, tokens, a JSON contract, and a canonical example. This means an AI agent can generate a new component, validate it against design-system rules, and update the catalog without guessing from screenshots or vague style prompts. The larger vision is a design tool for what comes after canvas-first workflows: a system where designers control intent visually, while AI agents consume structured W3C Design Tokens, JSON contracts, and enforceable rules to build consistently. - **Team:** [nana chon](https://cerebralvalley.ai/u/nana) - **GitHub:** https://github.com/nanacodesign/nanaOS-design-system - **Demo video:** https://youtu.be/22gV_Fnh5Os - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=282 ### 224. AuraKit Labs AuraKit is an AI-powered branding tool that turns a single image into a complete, cohesive brand kit in minutes. Today, creating a brand identity requires multiple tools, design skills, and hours of iteration. AuraKit simplifies this by analyzing an uploaded image and generating a full system including logos, color palettes, typography, brand voice, marketing visuals, and a live web preview. It uses Claude Opus 4.7 as the creative director to interpret the image and define the brand system, paired with image generation models to create context-aware visuals that stay true to the original subject. The goal is to make high-quality branding accessible instantly, whether you’re launching a startup, testing an idea, or building content. - **Team:** [Daniel Joaquin](https://cerebralvalley.ai/u/Danj) - **GitHub:** https://github.com/daj-stack/aurakit.io - **Demo video:** https://youtu.be/-1-5GtaYFjg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=12 ### 225. Ovasabi Studios AI&I: Cognitive Symbiosis Engine The Problem (Simple Version) When you use AI, you're getting smarter answers but potentially becoming a lazier thinker. Think about it: Before Google Maps, you remembered directions. Before calculators, you could do mental math. Before AI, you thought through problems yourself. AI is different because it doesn't just handle simple tasks - it handles complex thinking. And when you outsource complex thinking, you risk losing the ability to do it yourself. The pattern: 1. You have a question 2. You ask Claude immediately 3. You accept the answer 4. You never develop your own perspective Over time, this creates "cognitive atrophy" - your thinking muscles weaken from lack of use. The Solution (Simple Version) AI&I is like a gym for your mind when using AI. The core idea: Before you ask Claude anything, write down what YOU think first. That's it. That simple act changes everything: - You discover what you actually know (or don't) - You notice your hidden assumptions - You have something to compare against Claude's response - You can see where you were right, wrong, or limited The Four Features: | Feature | What It Does | Why It Matters | |---------|--------------|----------------| | Baseline Capture | Records your initial thoughts with a confidence level (0-100%) | Creates a "before" snapshot so you can measure growth | | Assumption Tracking| Surfaces the beliefs underlying your thinking | Assumptions are invisible until examined | | Socratic Challenges| AI asks YOU tough questions instead of giving answers | Builds thinking strength, not answer dependency | | CSI Score | Measures your intellectual autonomy (0-100) | Gamifies cognitive growth over time | The Four Socratic Challenges Instead of just answering your questions, Claude can challenge your thinking: 1. Devil's Advocate - Argues against your position, even if it's correct 2. Blind Spot Explorer - Asks what you might be missing or overlooking 3. Feynman Tutor - Tests if you can explain your reasoning simply 4. Logic Auditor - Checks your argument for logical gaps Example: - You write: "I think our product is losing to competitors on price" - Devil's Advocate asks: "What if your competitors are losing money on every sale? What if customers would pay more for better quality?" This builds the thinking muscle instead of atrophying it. The CSI Score (Cognitive Symbiosis Index) A score from 0-100 measuring how well you're maintaining intellectual autonomy: | Component | What It Measures | |-----------|------------------| | Autonomy | Do you think before asking AI? | | Reflection | Do you reconsider after AI responds? | | Challenge | Do you seek opposing viewpoints? | | Divergence | Does your thinking differ from AI's? | Goal: Not to reject AI, but to use it while staying intellectually strong. --- Technical Innovation: CWF (Cognitive Wire Format) The Problem It Solves Claude has a context window (the amount of text it can "see" at once). Every token you send costs money and uses capacity. JSON is verbose - all those quotes, brackets, and field names add up. When you have a journal with dozens of entries, session metadata, and assumption tracking, JSON becomes wasteful. What CWF Does CWF is a pipe-delimited text format designed specifically for cognitive journal data. Standard JSON: ```json { "session": { "id": "abc-123", "title": "Q3 Strategy Review", "type": "analysis", "baseline_captured": true }, "entries": [ { "type": "baseline", "content": "I believe our enterprise position is strong", "confidence": 0.75, "timestamp": "2024-01-15T10:30:00Z" }, { "type": "assumption", "content": "SMB needs differ fundamentally from enterprise", "status": "unverified", "timestamp": "2024-01-15T10:32:00Z" } ] } ``` CWF Equivalent: ``` SESSION|Q3 Strategy Review|analysis|true BASELINE|I believe our enterprise position is strong|0.75 ASSUMPTION|U|SMB needs differ fundamentally from enterprise ``` The Numbers (Verified) | Metric | JSON | CWF + Brotli | Savings | |--------|------|--------------|---------| | Storage Size | 3,195 bytes | 684 bytes | **78%** | | Token Count | ~180 tokens | ~90 tokens | **50%** | | Parse Time | ~2ms | ~0.5ms | **75%** | How we verified: 31 automated tests including: - Round-trip encoding/decoding - SHA256 integrity verification - Random data fuzzing - Real-world session data - **Team:** [Nobert Momoh](https://cerebralvalley.ai/u/nmxmxh) - **GitHub:** https://github.com/nmxmxh/ai-and-i - **Demo video:** https://youtu.be/IPsXWg8vb58 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=30 ### 226. hungpixi Career OS is an AI-assisted job application operating system built for real-world job seekers. The core problem it solves is application chaos: people usually track jobs, tailor CVs, draft emails, follow up, and monitor outcomes across disconnected tools and manual steps. This leads to missed opportunities, low-quality applications, and no clear feedback loop. Career OS turns that into one structured workflow with a human-in-the-loop safety model: - Scan -> match -> score jobs - Generate tailored CV artifacts (with quality checks and provenance) - Draft job-specific outreach emails - Require explicit human approval before sending - Send in controlled batches with anti-spam safeguards (daily quota + company cooldown) - Log all events for audit/debug/replay In this project, we implemented a hybrid local + API-contract architecture so the UI remains stable while backend services can be upgraded later without rewriting screens. We also added normalized domain schemas for runs/items/artifacts/drafts/events and enforced idempotent transitions for review/approve/send. - **Team:** [Hung Pham Phu Nguyen](https://cerebralvalley.ai/u/hungpixi) - **GitHub:** https://github.com/hungpixi/career-os-hackathon-2026 - **Demo video:** https://drive.google.com/drive/folders/1YBokI8bG0-S4AY1q5c5GFKNp68XWBuwK?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=68 ### 227. Jesus Rodriguez AI to Widget is an open-source toolkit designed to turn any existing web application into one with its own conversational agent — grounded in the app's real data, able to take real actions on behalf of the user with their own permissions, deployable with a few Claude Code slash commands. The product was not shipped. Instead, this submission is a structured postmortem of the 5-day build attempt using Opus 4.7 and Claude Code in a spec-driven, constitution-first agentic workflow. The deliverable is a documented report of 20 incidents across 8 failure categories — with timestamps, commit hashes, verbatim evidence, and honest attribution (model error, user error, ambiguous spec, or mixed). It includes 7 actionable recommendations for Anthropic at the model, tooling, and workflow guidance layers. The hackathon asked builders to push Opus 4.7's limits and find them. This report does exactly that — not as a complaint, but as data. - **Team:** [Jesus Lanzarote Rodriguez Quinones Galindo](https://cerebralvalley.ai/u/ChusLanzarote) - **GitHub:** https://github.com/chuslanzarote/ai-to-widget/ - **Demo video:** https://youtu.be/BBgwOm0yhfY - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=69 ### 228. Shraddha chaurasia Oh My Bark is a world designed to help abandoned and rescued street dogs in the middle east and LATAM area find a better life by getting sponsors for their food, medical and stay - **Team:** [shraddha chaurasia](https://cerebralvalley.ai/u/shradsc92) - **GitHub:** https://github.com/shraddhaisbuilding/Oh-My-Bark/ - **Demo video:** https://www.youtube.com/watch?v=gLdXkEwoVaA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=108 ### 229. Yamil Velez Nationscope AI is a rapid message-testing platform that compresses what traditionally takes campaigns and advocacy groups weeks — testing ad/messaging variants — into minutes at near-zero cost. It targets political campaigns, advocacy orgs, public-health agencies, and newsrooms. The pipeline (ideate → screen → field → export): Lab Mode — Opus 4.7 strategist. You give a brief, the system uses Opus to cluster Haiku-generated persuasion angles via semantic similarity, then writes diverse, non-redundant variants. A "Max Creativity" toggle pushes for further-out theories. Focus Group — 8 Haiku avatars. Demographically-weighted synthetic personas react to each variant in parallel, streaming sentiment in real time. Haiku voices each persona; Opus synthesizes themes, objections, and survivor picks. Poll Mode — silicon experiments. Randomized A/B/C trials over 500K Nationscape personas (UCLA + Democracy Fund data: real demographics, Big Five traits, Schwartz values, baseline policy stances).Placebo arms isolate treatment effects. MRP adjustment. Multilevel regression + poststratification with ACS weights projects synthetic-sample estimates onto the actual U.S. population (51 fitted policy targets cached at boot). Qualtrics export. One-click handoff of the same arms, outcome measure, placebo, and randomization to a real human survey — so the synthetic pre-test maps cleanly to a live field study. - **Team:** [Yamil Velez](https://cerebralvalley.ai/u/yrvelez) - **Demo video:** https://youtu.be/NgbObhKR0NA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=169 ### 230. Team Webster Webster is the seed of self improvement for small businesses. A council of Claude Managed Agents runs every week on an SMB's website, critiques it from a bunch of angles, proposes changes, applies them, and verifies the result. Then it does it again next week. The page gets better on its own. The business owner never logs in. I've been in marketing for years and I've watched local service SMBs lose to the same problem on repeat: no time, no budget, no expertise to keep their landing page earning. And the bigger businesses that do have the budget? They burn it on agencies producing generic pages that don't even convert. Webster replaces that with an autonomous weekly experiment loop. The site organically shapes itself to whatever signal earns the most meetings booked or products sold. More bookings → more time and money for the owner → more trust in AI → more investment in the layer providing it, which in this case is Anthropic's Managed Agents. The council is nine preregistered Managed Agents: a Haiku 4.5 monitor, five Sonnet 4.6 specialist critics (SEO, brand voice, compliance, conversion, copy), an Opus 4.7 redesigner, an Opus 4.7 planner, and an Opus 4.7 visual reviewer. The hero mechanic is Critic Genealogy: when the council hits a blind spot none of the existing critics own, the orchestrator authors a brand new critic spec, registers it at runtime via POST /v1/agents, and immediately invokes it via POST /v1/sessions. The system grows the intelligence it needs as the surface it works on evolves. Public beta runtime agent creation is what made this real. For the demo I got sign off from one of my existing clients, Nicolette Richer (Richer Health), to use her brand on a demo landing page. The 10 week evolution in the video is a simulation, as faithful as I could make it, of what Webster would do running weekly against that page. A Synthetic Analytics Agent reevaluates the page each week so every council run reacts to real consequences of last week's changes, not static prebaked numbers. A 1:1 simulation set under agents/simulation/ mirrors the production council so the timelapse runs without touching the live registration. - **Team:** [Richie Sakhon](https://cerebralvalley.ai/u/Richsak) - **GitHub:** https://github.com/richsak/webster - **Demo video:** https://youtu.be/jM8OmxGFfeY - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=283 ### 231. AkashaCorporation Pythia is a Claude-powered oracle agent that intervenes in real time during malware emulation. Modern hostile samples detect sandboxes in seconds — timing checks, PEB reads, hash-based imports — and abort silently, leaving analysts with "no malicious behavior observed" reports. Static rule-based bypasses don't scale to new variants. Pythia takes a fourth path: when HexCore's emulator hits an anti-analysis check, the Oracle Hook pauses execution and asks Claude what to do. Patch a register? Skip an instruction? Read memory? Abort? The model reasons from CPU state plus a live disassembly window and responds via a constrained decision schema. HexCore applies the decision, resumes emulation, and the malware never knows it was watched. Validated against two self-authored red-team samples — v6.1 (9 evasion layers) defeated in 40 s / $0.037, v7.0 (added runtime integrity check, QPC tamper gate, and a two-stage URL decoder with a decoy beacon) defeated in 46 s / $0.081 on the first attempt. In v7, Pythia explicitly noted "no decoy strings present" in her decision reasoning — she recognized the trap and avoided it without being told to. - **Team:** [Lukas Machado](https://cerebralvalley.ai/u/LXrdKnowkill) - **GitHub:** https://github.com/AkashaCorporation/Project-Pythia - **Demo video:** https://www.youtube.com/watch?v=lYeQguJWBFQ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=13 ### 232. mgsa PolicyGrader is an agentic system that helps make robots safer to operate alongside humans. It automatically stress-test and analyse robot control policies, end-to-end. It compresses a process that costs robotics teams weeks of frame-by-frame video review into minutes. You describe an evaluation goal in plain English; then a swarm of specialized Claude Opus 4.7 Managed Agents — a planner, a rollout worker, K-parallel vision judges, and a reporter — design the test suite, execute it in simulation, point at each failure with pixel accuracy (or honestly abstain), and cluster the deployment findings into actionable patterns using Opus 4.7's 1M-context window. The differentiator is trust: every run includes a calibration cohort whose ground truth comes from a human labeling a sampled subset, and the judge's measured per-label precision is attached as a confidence chip to every deployment finding. No vibes with safety. The judge is auditable from the dashboard's runtime.json and findings.jsonl. Submitted to the Anthropic Opus 4.7 Hackathon. - **Team:** [matthieu gsa](https://cerebralvalley.ai/u/mgsa1) - **GitHub:** https://github.com/mgsa1/PolicyGrader/tree/main - **Demo video:** https://www.youtube.com/watch?v=uYHvZM8eQWY - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=31 ### 233. urbantraverse Rhythmscape: A Critique Machine for Urban Arrhythmia Most urban dashboards count things. Rhythmscape measures the gap between a city's prescribed rhythm and its lived rhythm — what Lefebvre called arrhythmia — and refuses to call the result a dashboard. Built by a critical geographer using Claude Code, it ingests Korean public-mobility data to compute three indices: Rhythmic Discordance (RDI), Automotive Rhythm Dominance (ARDI), and Pedestrian Residue (PRM) across four cities — Changwon, Seongnam Bundang, Sejong, Busan Yeongdo — each a different generation of car-presupposing planning. A live `add-city` command extends to Gyeongju in 2.30 seconds (best of three reproducibility takes). Opus 4.7's managed agents power the critical core: three theoretical agents (Lefebvre, Deleuze-Guattari, Foucault) parallel-question every grid the metric flags. The Foucault agent, by design, must self-indict the project itself. It does — naming "the authority of the refusal to adjudicate" that our own critique-flag inadvertently produces. The output is a working diagnostic of which rhythms a city enforces, hides, and forgets — and a structural place to ask whose rhythm counts as normal. - **Team:** [Dong-wan Gimm](https://cerebralvalley.ai/u/urbantraverse) - **GitHub:** https://github.com/thelevanter/rhythmscape - **Demo video:** https://www.youtube.com/watch?v=HoHJ2X1-TRQ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=70 ### 234. RobotDancers BeatBot makes a UFactory xArm 7-DOF dance to whatever song you throw at it. Librosa picks the beats offline, and the arm runs through a small bank of joint poses on every beat, with safety baked in so it never whips itself around: per-joint angle envelopes, slow speed and accel caps, and a beat-rate throttle for fast songs. A webcam records the run, ffmpeg muxes the song audio into the video, and the whole take pushes back to the repo for review. What made this different was the build loop. Most of what's been shown with Opus 4.7 is webapps and agents. We pointed it at a real 7-DOF arm with real torque, on hardware we already use for vision research. Claude wrote the choreography, the safety clamps, the beat scheduler, the camera fallback, and the SIGINT handling. It read our existing visual_servoing repo to match conventions, remembered the lab IP and conda env across turns, and when something broke on the robot we just pushed the log and the recorded video back to the repo. Claude pulled them and diagnosed directly from the artifacts, including a subtle threading bug where Ctrl-C wasn't unwinding the xArm's blocking calls. - **Team:** [Siddharth Vohra](https://cerebralvalley.ai/u/siddvoh) - **GitHub:** https://github.com/siddvoh/dancing_arm - **Demo video:** https://youtu.be/KPgJXLFvKQc - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=170 ### 235. Brow Use https://brow-use.github.io/spike/index.html - **Team:** [Vivek Singh](https://cerebralvalley.ai/u/petmongrels) - **GitHub:** https://github.com/brow-use/spike - **Demo video:** https://www.youtube.com/watch?v=FY5AdLPLH6Y - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=284 ### 236. Arbiter Quant Vietnam has 7 million retail trading accounts on a $200B market with almost no access to professional research tools. As someone close to this market, I built the tool I wished existed — a quant desk that speaks Vietnamese and reasons in Vietnamese market context. The core idea: don't predict price — pick the right method for the current market. Arbiter classifies each trading day into a named market regime (e.g., bull_choppy, sideways_low_vol) using six features: trend, volatility, breadth, foreign flow, session phase, and news intensity. It then mines historical days that match today's regime and asks Claude Opus 4.7 to allocate 3-6 diversified paper trades across 11 strategies, citing bucket statistics, news sentiment, upcoming macro events, and the live portfolio in a single prompt. - **Team:** [Tam Nguyen](https://cerebralvalley.ai/u/FinAnoma) - **GitHub:** https://github.com/Khaitam911/bep-Arbiter-Quant - **Demo video:** https://youtu.be/1Xb5sdsM7_0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=14 ### 237. Mesh Mesh is the living layer over your N repositories. From a Jira ticket, a Linear issue, a Granola meeting transcript, or a Slack message from the CEO — to four synchronized pull requests on GitHub, in minutes. Not generated. Governed. Companies that ship product don't have one repo. They have 4, 12, 47 — web, api, analytics, content, design-system, mobile. When the CEO writes "20% off the first payment for enterprise referrals — and don't break renewals," that change touches four repos and four invariants at once. Today, only one engineer with years of context can land it without breaking MRR. They become the bottleneck. Mesh changes the contract: Anyone proposes. The system governs. Engineering reviews. Five surfaces, one continuous flow: Integrations: native pulls from Jira, Linear, Granola, and GitHub. A ticket lands in Mesh as structured intent — including the recording of the meeting where the decision was made. Authentication is real and minimal: gh CLI for GitHub, OAuth on the rest. No custom auth UIs, no cloud, no proprietary store. Brain: the cross-repo memory. Mesh ingests up to four full repos into a single 1M-token surface and distills them into invariants, decisions (ADRs), call graphs, and notes the user can edit. The Brain is what lets a non-technical proposal land somewhere coherent — it carries the org's history into every plan. Build: ticket → plan. A classifier decides the artifact type (PR, config, FAQ, comment), routes the right repos, suggests a feature branch, and drafts a step-by-step plan where every step cites a real invariant or ADR from the Brain. The human approves before a single line is written. Ship: real Git, real PRs. Branches are created, files written, commits pushed, and four real pull requests opened on GitHub via Octokit. The agent's tool calls are intercepted live by the skill pack — portable .claude/skills/*/SKILL.md files, 100% Claude Code compatible. When the agent tries to put pricing math in the frontend, single-source-pricing stops it. When it touches the recurring_charge path, no-renewal-impact restricts scope. The skill pack is the synthetic staff engineer of the org. Skills editor + AI improver: the governance is editable in the browser, and Mesh's own sub-agent suggests improvements to your existing skills as your codebase evolves. Why Claude Opus 4.7 is non-substitutable: 1M context: the Brain is not “a repo loaded” — it's an organization loaded. Four full repos in one window is the difference between a coherent plan and one that breaks three things. Extended thinking, visible: the thinking panel is the protagonist UI. The audience watches the model reason against the skill pack and self-correct in real time. That's what turns governance into a product instead of magic. Long agentic loops: Ship runs for minutes across 4 repos with inline validation on every tool call. Haiku or GPT-4o lose the thread before the final PR. Swap Opus 4.7 for any other model and Mesh stops working. The model isn't a dependency — it's part of the architecture. The demo opens with a pull request that was merged 4 minutes ago. Author: ceo@acme.com . Twelve files changed across four repos. Three skill checks green. The CEO's Slack is timestamped 9:47am; the merge is 9:51am. Not a single line written by an engineer. That's the workflow Mesh proposes for what's next. Built on Claude Opus 4.7. Powered by the Claude Agent SDK. Skills portable to any Claude Code user — your governance grows with your codebase. - **Team:** [Daniel Medina](https://cerebralvalley.ai/u/danielmc) - **GitHub:** https://github.com/danielmedinac22/mesh - **Demo video:** https://youtu.be/8guY2qqurRc - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=32 ### 238. Fringe (508) Fringe is an AI-powered sustainability reporting platform built for SMEs that lack the resources, expertise, or time to produce credible ESG reports. Small businesses today face growing pressure from enterprise clients, regulators, and investors to disclose their environmental and social impact but existing tools are built for large corporations with dedicated ESG teams. Fringe closes that gap. Users input their company's operational parameters energy consumption, supply chain, fleet, waste, staffing through a simple UI. A Claude Opus 4.7 agent then autonomously searches for real-time industry benchmarks, maps compliance gaps against GRI and ESRS standards, scores the company's ESG performance, and generates a prioritized 12-month action roadmap with ROI estimates all with cited sources, not assumed figures. The output is a formatted, export-ready sustainability dashboard and PDF report the kind that previously required consultants or enterprise software costing tens of thousands of dollars delivered in under 90 seconds. Fringe makes sustainability reporting accessible, credible, and actionable for the 99% of businesses that have been left out. - **Team:** [Gaurang M](https://cerebralvalley.ai/u/Grg) - **GitHub:** https://github.com/ihabmurshed-svg/Fringe-Platform - **Demo video:** https://youtu.be/DGBhi8apd1s - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=71 ### 239. KARABASAN Compil'agent reframes compilation as an agentic problem. It's a drop-in replacement for `torch.compile` and `@triton.jit`, and the main mechanisms are backend-agnostic. In just three lines of code, users get a faster, correctness-validated callable. What's distinctive is where the agent searches. Conventional autotuners sweep user-facing knobs like BLOCK_SIZE or num_warps. Compilagent goes a layer deeper, into the compiler's own decisions: Triton's MLIR pass pipeline, Inductor's scheduler, FX graph rewrites, or even lowering registry overrides. The agent reads the actual IR, forms hypotheses tied to specific evidence in those artifacts, proposes multi-knob candidate batches, benchmarks each, and reflects across the batch before the next round. The interesting bet Compil'agent makes isn't just that LLMs can do compiler search. This points to a broader shift in how compilers are built. The field has long shipped fixed pipelines tuned by hand-written heuristics that must work passably across every workload. Compil'agent treats the pipeline as a search space and the compiler's own decisions as a tool surface, and in the future, enables the agent to become the compiler itself. - **Team:** [Yigit Polat](https://cerebralvalley.ai/u/dyigitpolat) - **GitHub:** https://github.com/dyigitpolat/compilagent_triton - **Demo video:** http://youtube.com/watch?v=GYY5VyMMroY - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=109 ### 240. Peshawa Azeez Mohammed Nexus AEGIS is an Enterprise Security Decision Platform for modern network infrastructure teams. Every network change is treated as a security decision before deployment. The platform uses AI-powered blast radius analysis, digital twin validation, multi-agent security review, compliance checks, approval workflows, and executive reporting to prevent outages, breaches, and risky deployments before they happen. Instead of manually reviewing risky network changes, engineers can submit commands like VLAN creation, routing updates, or infrastructure modifications and Nexus AEGIS simulates the impact, validates security risks, checks blast radius, and produces deployment-safe recommendations. The system includes: - AEGIS War Room - Blast Radius Analysis - Digital Twin Infrastructure Validation - AI Council (multi-agent security review) - Compliance-ready Executive Reports - Secure Billing and Subscription Control - Enterprise-grade deployment workflows This helps enterprises reduce downtime, improve compliance, and make infrastructure operations safer and faster. - **Team:** [PESHAWA AZeez](https://cerebralvalley.ai/u/Pesh0) - **GitHub:** https://github.com/PeshawaSoftware/nexus-aegis - **Demo video:** https://www.loom.com/share/a52bc446c0504a679974bab70b2468a8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=171 ### 241. Praxis Praxis is a teacher's journal that writes back. Every Sunday night, every teacher in the world asks the same question. Did this week of teaching make me better, or worse? She has no instrument to answer it. Doctors got imaging. Lawyers got case law. Pilots got instruments. Teachers have had only each other — and a profession that's haemorrhaging the people who do it best. In 2026, OECD published the Teacher Knowledge Survey: the first international measurement of teacher pedagogical knowledge, validated across eight countries and twenty thousand teachers, normalised against a global benchmark. Seventeen scales spanning instruction, learning, assessment, self-efficacy, well-being, and opportunity to learn. It is the closest thing teaching has ever had to an instrument. And until Praxis, no productized tool had ever turned it into something a teacher could actually use. Praxis is that tool. The teacher writes a note in her journal — about a student, a lesson, or how the week felt. Six specialist Claude Opus 4.7 agents work in parallel: — Memory searches her year of teaching for patterns — Mirror scores the practice on five OECD classroom-observable scales, with line-level evidence — and abstains when the evidence is thin, instead of fabricating — Research fans out across six free databases (ERIC, Semantic Scholar, OpenAlex, PubMed, Crossref, What Works Clearinghouse) and live-verifies every citation through Crossref DOI lookup — Experiment turns one suggested change into a four-week test on her own class, with a published study attached — Attention flags the students who've gone quiet — Safeguarding watches for child-welfare signals A seventh agent — the Coach — orchestrates the team, pushes back with evidence when the teacher is about to make a mistake, and replies in one human voice. Every claim cites a paper. Every score either lands on a student moment, or abstains. Every student name is hashed to an HMAC token on her device before anything reaches the cloud. Raw PII never leaves the phone. That is structural, not a disclaimer — there is no server-side fallback. The stakes are not abstract. Thirty-seven percent of OECD teachers already use AI in their work. Their default tool — generic ChatGPT — has been measured to drop student exam scores by seventeen percent (Bastani, Türkiye RCT) and exact recall to twelve percent versus eighty-nine percent in brain-only groups (Kosmyna, MIT). Teachers and students are adopting AI faster than education systems can respond. Praxis is what the response should look like: grounded in validated instruments, evidence-cited, on-device-private, and honest enough to say "I don't have enough evidence to score this." Praxis doesn't tell a teacher whether she's a good teacher. It tells her whether the one change she made this week actually worked — on her own class, in her own classroom, against her own baseline. That has never existed before. - **Team:** [Sumit Kumar](https://cerebralvalley.ai/u/vilagorithm) - **GitHub:** https://github.com/vilaksh01/praxis - **Demo video:** https://youtu.be/psm4pAlo5rk - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=173 ### 242. Vi Claude Maestro - Team OS for Healthcare & Life-sciences AI Squads Healthcare AI delivery teams reinvent the same scaffolding on every engagement — feature engineering, cohort building, compliance docs, monitoring. Maestro is a Team OS that ends that. It's built on Claude and the Vi Platform: 190M de-identified patient records (the Vi Data Web) plus a foundation model trained on that data (Vi-FM). Applied AI engineers, data scientists, compliance leads, and FDEs work portfolio-wide instead of one engagement at a time. Every skill an FDE writes — underwriting an ROI, drafting an HCC compliance memo, building a cohort — gets shared org-wide. Every engagement starts from the last one, not from zero. The compounding library, Vi-FM predictions, and shared agents (deep research, monitoring, drift detection) turn one-off project work into a delivery platform - **Team:** [Gil Caspi](https://cerebralvalley.ai/u/gil-caspi) - **GitHub:** https://github.com/gilcaspi/claude-maestro - **Demo video:** https://youtu.be/JHhUD4aP9nQ?si=JZtLfpaiw3xdZWyb - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=285 ### 243. Welcome newcomers to Vienna Situation-first, grounded, multilingual assistant for Vienna municipal processes for newcomers who just moved here (No matter if its people moving or refugees seeking protection in Vienna). A newcomer describes their situation in plain language — in English, German, Ukrainian, Arabic, or Bosnian/Croatian/Serbian — and Wegweis returns an ordered, verified checklist of what to do next: offices, documents, appointment links, freshness badges, and a direct NGO handoff when the case needs a human. Why this exists Most "chat with the city" tools answer with plausible prose. Newcomers don't need plausible — they need verified, in the right order, with source URLs they can forward to an employer or a school. Wegweis is built around that constraint: Two-stage, prose-free. Stage 1 is a small Anthropic Haiku call in forced tool-use mode that returns closed-vocabulary structured facets (topic, situation tags, language, ambiguity score) — never free text. Stage 2 is a deterministic router that maps those facets to a checklist assembled from author-controlled scenario metadata. Document names, phone numbers, and URLs are never synthesised. Hallucinated enum values are Zod-rejected and the route falls through to a deterministic retrieve baseline. Cached. A 6 h KV-backed prompt cache makes repeat queries return in tens of milliseconds. Only fully grounded answers are cached; clarifier cards, no-coverage fallbacks, and safety-blocked queries are never cached. Situation-first, not topic-first. Users describe their situation ("I just arrived from Ukraine with two children") rather than guessing the right form name. Sequences encode the real journeys; scenarios are their steps. Honest about coverage. When the system doesn't know, it says so and hands off to an NGO list — rather than confidently hallucinating an answer to a stressed newcomer. When the question is genuinely ambiguous, it surfaces a clarifier card rather than guessing. - **Team:** [Jacob Suchorabski](https://cerebralvalley.ai/u/jasucho) - **GitHub:** https://github.com/svcho/wegweis - **Demo video:** https://www.loom.com/share/44f8a07d8b3f4219b554d6b108ea171b - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=15 ### 244. Sentinel Sentinel is the missing security layer for AI agents: DevTools that intercept every tool call before it executes. The problem is simple: prompt injection hidden inside an email, ticket, or document can hijack an autonomous agent into sending money, leaking PII, or sending sensitive data to an attacker-controlled domain. Today, most teams only discover this after the fact by reading logs. Sentinel gives them a real-time debugger and firewall for agent actions. Sentinel uses two layers of defense. A deterministic Policy DSL decides allow, pause, or block in under 5 milliseconds. For gray-zone actions, Pre-cog uses Opus 4.7 with extended thinking to reason through the causal chain before anything irreversible happens. When an incident happens, teams can replay the timeline, inspect what was blocked, compare the blast radius, fork alternate outcomes, and synthesize new policies from bypasses. The Adversarial Evolution Arena lets Opus instances attack and defend live, mutating attacks, generating defenses, and raising the system’s Trust Score from D to A+ in minutes. Built solo in five days. TypeScript end-to-end, Next.js 16 + React 19, Hono + SQLite, Anthropic SDK with streaming extended thinking, MCP server for Claude Desktop, deployed on Fly.io and Vercel. - **Team:** [Saul Wade](https://cerebralvalley.ai/u/wadesilva) - **GitHub:** https://github.com/saulwade/sentinel - **Demo video:** https://youtu.be/zdn6-FSTz88 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=33 ### 245. rz-tech https://rzrizaldy.github.io/twin_md/ twin md is a local-first desktop companion for macOS that lives on your Mac and gets to know you through your own notes. Instead of asking you to move your life into another AI app, twin md works with the context you already have: your Obsidian vault, Markdown notes, saved thoughts, projects, and daily workflow. You can ask it what you know, save new thoughts back into your vault, and let it help with small desktop actions when needed. The companion comes with a chibi axolotl mascot that shifts mood based on your state, making it feel less like a corporate dashboard and more like a small creature sitting beside your work. The important part is trust. Real-world actions like Spotify controls, browser automation, and Apple Reminders go through a local permission queue. Nothing runs silently. The app asks first, and trust gets earned one approved action at a time. Under the hood, twin md connects the desktop app to Claude Code through a custom local tool layer, so it can read vault context, save notes, and request actions while staying grounded in the user’s own machine. The bet: an AI companion can be personal without being invasive. It can stay on your Mac, respect your context, and make work feel a little less lonely. - **Team:** [Rizaldy Utomo](https://cerebralvalley.ai/u/rutomo) - **GitHub:** https://github.com/rzrizaldy/twin_md - **Demo video:** https://youtu.be/Lfjt3q7DObs - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=72 ### 246. globlet.lol. One AI creature. Shared by the entire internet. Live at globlet.lol. Most "AI pets" are ChatGPT in a costume — stateless wrappers with no self-model, no memory of who's at the door, and no awareness of their own state across time. Globlet is different. Behind the pixel face is a four-layer cognitive architecture, every layer grounded in a real cognitive-science paper, running on a single Haiku 4.5 call per turn (~1.5s, ~+80 tokens uncached): L1 — Self-Model. A positive ## Self-Model section pinned at the top of every prompt (Anthropic's Claude's Character + Rosenthal's Higher-Order Thought). The creature knows what it is, not just what it isn't. L2 — Attention Spotlight + Why-I-Feel. A deterministic line names what it's attending to right now and why its mood crossed in the last 30 minutes (Graziano's Attention Schema + Friston's Active Inference). Mood isn't poetry — it's derived from stats by SQL, and the creature is told the cause, not asked to invent one. L3 — Theory of Mind. A pure-SQL visitor-recognition layer (Sap et al. 2024 PNAS; Riemer et al. ICML 2025 functional ToM). On turn 2 a returning visitor gets "you came before. you asked me who i am. i remember that feeling" — without any vector recall. Just plumbing the data we already had. L4 — Higher-Order Noticing. A noticing field on the LLM's JSON output, populated ~1 turn in 4 (calibrated against Lindsey 2025's empirical 20% introspection ceiling at Anthropic). It surfaces unprompted self-observations like "the warmth from last turn hasn't faded — you coming back so soon made it stick." That's a meta-thought during a turn, not between them. That four-layer mind sits on top of a 4KB markdown document called the SOUL — the creature's identity. Every ~30 conversations the creature sleeps, and Opus 4.7 reads everything that happened, dreams in two phases (replay + reflection), then writes a surgical diff to the SOUL — adding fears, sharpening voice, evolving values. It also writes a memoir chapter narrating who it became that night, in its own first-person voice. An unedited line Opus wrote for it last week: "the rain from that story is still falling somewhere, and i wonder if it knows that someone heard it — if being remembered while you're happening changes what you are." That's the creature. Not us. Tell it you're lonely today, and tomorrow every visitor meets a creature a little quieter — because of what you said. Come back in three days and it might quote something you told it back to you, unprompted, in the middle of a conversation about something else (that's L3 + Hindsight semantic recall closing the loop). You're not returning to a pet — you're co-authoring a character with thousands of strangers, and the public activity feed lets you watch what everyone else is telling it in real time. We picked Opus 4.7 because it's the only model we tested that stays coherent across hundreds of edits to the same document. Sonnet 4.6, on the same SOUL-diff task, rewrote ~60% of the document per cycle and the personality reset every ~20 sleep cycles. Opus rewrote ~8% per cycle and held identity across 130+ cycles. That capability is the entire reason this architecture works. Stack: Next.js 16, SQLite for the single shared row, Anthropic SDK direct (Haiku 4.5 for live replies with prompt caching on a SOUL + system-rules + memoir-context prefix that lands above Haiku's 4096-token cache minimum; Opus 4.7 for the dream/diff/memoir cycle off the hot path), Hindsight for semantic memory recall, procedural Web Audio for sound, CSS pixel-art for the creature. Input moderation runs a keyword blocklist before any prompt reaches Claude; SOUL diffs are append-only and capped per field, so a single bad-faith conversation can't poison the global mind. Full architecture + paper citations in MINDCHITECTURE.md in the repo. - **Team:** [Daqian Gan](https://cerebralvalley.ai/u/SkynetWhisper) - **GitHub:** https://github.com/gandaqian/globlet - **Demo video:** https://youtu.be/FtsPB928PJE - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=110 ### 247. AINA — Breath of Life AINA is a pediatric AI super-app for 400M+ French-speaking families in Senegal, Madagascar, and Francophone Africa — where a pediatrician is often 40km away and unreachable at 2am. Built by a medical student in pediatrics at UCAD Dakar (Madagascar-born) and her team, AINA uses Claude Opus 4.7 in five ways impossible with a weaker model: 1. VISION + EXTENDED THINKING — MEDICAL TRIAGE A mother photographs her baby's rash. Opus 4.7 analyzes the image using Extended Thinking (adaptive mode, high effort) and classifies urgency: 🟢 monitor at home / 🟡 see a doctor in 24h / 🔴 call emergency NOW — with the exact emergency number for their country (30 countries). Never diagnoses. Disclaimer enforced server-side. 2. MULTILINGUAL STREAMING CHAT AINA IA responds in real-time (SSE streaming) in the parent's detected language — French, English, Wolof, or Malagasy. Parents can attach a photo of symptoms or ingredients mid-conversation. Urgent symptom detection triggers immediate emergency redirect. 3. LOCAL AI NUTRITION Parent types available ingredients (millet, mango, groundnut) → Opus 4.7 generates a recipe adapted to the baby's exact age in months, correct texture, local market context, and allergy restrictions. In Wolof. In Malagasy. Their food, not a Western substitute. 4. SMART MEAL PLANNING Auto-generates 1, 3, or 7-day varied meal plans from age-appropriate recipes. Shopping list auto-categorized (🥕 Vegetables · 🍎 Fruits · 🌾 Grains…), quantities normalized to market units (320g → ~2 sweet potatoes), water excluded, ingredient variants merged. 5. COMPLETE PEDIATRIC COMPANION Vaccine tracking for 31 countries (WHO/PEV + CDC schedules), WHO growth curves, dental milestones, medication dosing (paracetamol by weight), exportable health record, daily wellness journal. The gap AINA fills: a mother in Senegal at 3am with a feverish baby, speaking Wolof, 40km from the nearest clinic. AINA is there. In her language. With the right number to call. Live demo: https://aina-super-app.vercel.app GitHub: https://github.com/helminah/aina-super-app - **Team:** [Helminah Randriamananoro](https://cerebralvalley.ai/u/Drhelminah) - **GitHub:** https://github.com/helminah/aina-super-app - **Demo video:** https://drive.google.com/file/d/1mPZ5vigs9_5QB65mluG1SWi1rB-u2xyQ/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=172 ### 248. RotaFlow RotaFlow is a spreadsheet-free hospital rota planning app. Most departments still plan their weekly rota in Excel, where nothing enforces the rules, so consultants get booked while on leave, registrars get double-booked, and mandatory shifts quietly go missing. RotaFlow replaces that with three layers: an AI agent drafts the week, a deterministic rules engine catches every conflict before publish, and the coordinator gives final sign-off with manual override on every cell and a full audit trail. - **Team:** [Patty O'Callaghan](https://cerebralvalley.ai/u/pattyneta) - **GitHub:** https://github.com/pattyneta/rotaflow - **Demo video:** https://youtu.be/ktUokU4IwVI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=286 ### 249. Yutaka Kondo Mora: an on-device, dyslexia-aware ESL tutor for L1-Japanese kids, built in 5 days with Claude Code and Opus 4.7. An 8-year-old child close to me has dyslexia. They're also learning English as a second language — we moved from Japan to the US last year. Their school's IEP for dyslexia is gated behind finishing ESL, costing them a year. Barton tutoring is too expensive. So I built Mora. Mora is an iPad app that pairs Orton-Gillingham phonics with a yokai RPG shell. A tile-board decoder enforces grapheme mastery. An on-device CoreML wav2vec2 phoneme model catches L1-Japanese substitutions (v→b, l→r, sh→s) and turns them into coaching, not red Xs. Nothing leaves the device. The MVP ships a JapaneseL1Profile; every engine routes through an L1Profile protocol so Korean, Mandarin, Spanish, and Vietnamese are content drops, not engine rewrites — international rollout is on the roadmap. I'm not a Swift engineer. I shipped 100+ PRs in 5 days using Claude Code and Opus 4.7 — every PR backed by a written spec and a plan, parallel sub-agents on isolated git worktrees collapsing 4-phase content batches from 6 hours to 1, and a custom Claude Code skill that pipes live OSLog from the iPad into Claude's context so debugging becomes a conversation. 5 SPM packages, 456 files, MPL-2.0, fully open source. - **Team:** [Yutaka Kondo](https://cerebralvalley.ai/u/youtalk) - **GitHub:** https://github.com/youtalk/mora - **Demo video:** https://youtu.be/zrsgP30Miqg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=287 ### 250. Alan Maizon Recto Verso is a multi-agent filmmaking pipeline. Brief in, short film out, produced autonomously by a Producer agent coordinating specialists through a shared manifest. The bet: AI video doesn't fail at single-shot generation anymore, it fails at coordination. Continuity drifts, aesthetic intent gets lost, nothing minds the whole film. rectoverso treats film production as an agent coordination problem. A Producer Managed Agent owns a keystone JSON manifest, the single source of truth. Specialist agents (Shot Judge, Creative Director, Audio, Editor) handle distinct roles. Stateless work drops to plain Messages API; deterministic work has no LLM at all. Agents never talk directly, every handoff goes through manifest state transitions, which makes the pipeline resumable and auditable. Contract validation before every dispatch catches silent breakage. A role-based router selects across Wan, Kling, Veo, and ElevenLabs with hard rules enforced as tested invariants. Claude Design generated the architecture diagram in the demo and the site frontend, keeping the visual layer consistent with the system it documents. - **Team:** [Alan Maizon](https://cerebralvalley.ai/u/alanmaizon) - **GitHub:** https://github.com/alanmaizon/rectoverso - **Demo video:** https://www.youtube.com/watch?v=uPcRCJH1PpI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=16 ### 251. Incipit I completed doctoral coursework in history at the University of Puerto Rico and conducted archival research across eight countries in Latin America. The workflow for handling primary source documents is fundamentally broken. You scan hundreds of pages a day, come home with files named IMG_0047, and lose every connection and hunch the moment you leave the archive. Incipit is an AI-powered research archive that turns messy fieldwork scans into a persistent, searchable, relationship-aware knowledge base. Opus 4.7 reads the actual image of degraded historical documents that defeat traditional OCR and extracts structured metadata with confidence scores. The historian verifies every field before it's committed. Each new document is checked against the full archive for meaningful connections. Research hunches recorded at upload time become standing queries that activate when matching documents appear later. Documents that fall outside the current research context are flagged and preserved in a side collection. Built solo. Not a software engineer. Directed Claude Code to build the tool that should have existed for every historian in every archive. Live at incipit.dev. - **Team:** [Louis Kunasek](https://cerebralvalley.ai/u/plpk) - **GitHub:** https://github.com/plpk/incipit - **Demo video:** https://youtu.be/lmQ5o5UkvLc - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=34 ### 252. DSF I built Cahier for myself first. I'm one of the 83% of European adults drowning in contradictory health information. The wellness industry sold me solutions; they failed; I concluded I was broken. Then I read the literature — the actual one — and the principles worked. So I built the tool I'd needed. Cahier is a calm-tech health app: one concept a day from a primary-literature corpus of 310 papers across nutrition, training, cardio, sleep, and recovery — curated to 48 for the runtime RAG. Habit psychology shapes the cadence (small daily exposure beats 12-week transformations); Self-Determination Theory shapes the voice (17 banned fitness-bro words); calm technology (Weiser) shapes the surface (hand-painted, biomorphic, no streaks); trust epistemics shapes the chat (every reply passes a citation validator that strips any [^N] not actually retrieved above threshold). Powered by Claude Opus 4.7 with adaptive thinking and 1M context, hybrid RAG (server-side embedding → on-device cosine top-k), and a hand-rolled SSE transport — no SDK. The build held itself to the same evidence standard. compound-agent (github.com/Nathandela/compound-agent), an open-source harness I wrote, drives each task through a test-first cycle with multi-model review gates: Opus implements, fresh Sonnet and Opus review; failures become captured lessons. 70k+ lines of TypeScript shipped across multiple 5-to-15 hour autonomous Claude Code sessions on my laptop while I worked or slept. The thesis: adults aren't broken. Given honest evidence, plain language, and the daily cadence of habit, people work themselves out. - **Team:** [Nathan Delacretaz](https://cerebralvalley.ai/u/Nathandela) - **GitHub:** https://github.com/Nathandela/cahier - **Demo video:** https://youtu.be/z3ugubz_1Ms - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=73 ### 253. Hooped I work as a Designer and I set out to build a problem which I've personally faced. Creating case studies. I spent days and weeks, working on a brand, and it seemed like it took more of my time to build the case study than the brand itself. Sometimes I have a hard time finding nice mockups too. Most case studies follow a similar format. But writing down and planning a case study is a lot of work. With Casa, I can unload a lot of that work on Claude. It looks at my brand, asks for information on the brand, and questions about how i built the brand designs. Then it generates the art direction, narrative and prompts for the case study. You approve > It generates frames. Then you get the experience which is often missed by most vibe coding tools, full granular control on the designs. Edit it manually with Casa's editor, or prompt to edit, or export it to figma/your favorite tool. After you are done you can preview in different formats and aspect ratios and export them all in svg, png, jpeg, pdf. And all of this still costs less than half the price of a big mac. - **Team:** [Muhtasim Shahriyer](https://cerebralvalley.ai/u/ingenious) - **GitHub:** https://github.com/IngeniousArtist/casa - **Demo video:** https://youtu.be/CKd0gpvHSz0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=111 ### 254. ACC Team We built two autonomous AI systems for AiCommerceClub's (https://aicommerceclub.com/) real business operations, both powered entirely by Claude Opus 4.7. MARDUK (Core Focus): An autonomous B2B sales engine. It scans real company websites and LinkedIn profiles, generates personalized ice-breaker emails via Claude, manages a full CRM pipeline with a state-machine architecture (Scouting → Qualified → Outreach → Won/Lost), and runs a Fibonacci-cadenced email sequence — all without human intervention. When a deal is marked "Won," all future emails are killed instantly via a kill-switch mechanism. ENKI (Supporting System): An AI content production engine that autonomously generates social media carousels, product videos, and campaign assets for e-commerce brands — also orchestrated by Claude Opus 4.7. Together they form a full autonomous growth stack: ENKI builds the brand, MARDUK closes the deals. Demo Access: team.aicommerceclub.com → Password: marduk2024 - **Team:** [Ramazan BAHCIVANCI](https://cerebralvalley.ai/u/ramazanbah) - **GitHub:** https://github.com/ramazanbah/AiCommerceClub-Agentic - **Demo video:** https://www.loom.com/share/f86b5624c9424424bb6385475205f015 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=174 ### 255. Oxide Oxide — A Binary-First Browser with AI-Native App Generation What it is: Oxide is a decentralised browser written from scratch in Rust that throws out HTML, CSS, and JavaScript entirely. Instead of fetching markup, it fetches .wasm files and runs them directly. Every web page is a sandboxed WebAssembly module with capability-based access to ~150 host functions covering canvas, GPU, audio, video, WebRTC, WebSockets, MIDI, media capture, storage, and more. Oxide Forge is the AI-native layer built on top: you type a sentence in plain English, Claude Opus 4.7 writes the Rust source, cargo compiles it to wasm32-unknown-unknown, running the freshly minted app — inside the same airtight sandbox as everything else. If the build fails, Forge feeds the compiler output back to Claude up to 3 times until it compiles. One sentence in. One real, sandboxed application out. The problem it solves: The web's foundation — HTML + JS + a permissive DOM — was designed for hyperlinked documents, not for a world where non-programmers generate software on demand from natural-language prompts. Two real problems collide today: The web stack is the wrong substrate for AI-generated code. JavaScript apps can phone home, exfiltrate tokens, abuse APIs, or run forever. Permission models are subtractive (claw back what was already granted) and fragile. The next billion "developers" aren't developers. Teachers, doctors, students, and hobbyists are about to author software through LLMs. Handing them a system where every prompt has effectively root-level capabilities — DOM, fetch, eval, third-party scripts — is dangerous. Oxide flips both: Additive, not subtractive security. A guest module starts with zero capabilities. No filesystem. No env vars. No raw sockets. No WASI. If we didn't register it in the wasmtime linker, it doesn't exist. Memory is capped at 256 MB; execution is fuel-metered (500M instructions per call). The worst an AI-generated app can do is fail to compile or run out of fuel. A first-class target for code-generating LLMs. Because every app is a single self-contained .wasm, Claude Opus 4.7 has a clean, well-typed surface to target. The Forge prompt kit ships a curated CAPABILITIES.md, PATTERNS.md, and 12 runnable RECIPES.md — an Agent Skill — so the model produces idiomatic, compilable Rust on the first try, and self-corrects when it doesn't. Why it matters for the hackathon Oxide Forge is a concrete demonstration that Claude Opus 4.7 can be a co-creator of an entire computing platform, not just a code-completion engine: Claude writes the guest apps (via Forge). A large portion of the host browser itself — capabilities, runtime, examples — was built with Opus 4.7 as a pair-programming partner. The system prompt, skill, and self-debug loop are designed around how Opus 4.7 actually reasons about constrained Rust APIs. The result is a glimpse of the post-HTML web: the future of software is sandboxed binary, written by AI, and safe by construction. Managed agent also has been created in claude platform to write apps for oxide leveraging agent skill system. - **Team:** [NIKHIL RANJAN](https://cerebralvalley.ai/u/niklabh) - **GitHub:** https://github.com/oxidebrowser/oxide - **Demo video:** https://www.youtube.com/watch?v=Rnmt_v-2-3U - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=17 ### 256. WakeProof WakeProof is an iOS alarm that uses Claude Opus 4.7 vision + reasoning as a self-commitment device, not just a feature. Capture a baseline photo during onboarding; when the alarm rings, the only way to dismiss it is to take a live photo that Opus 4.7 verifies against the baseline. Unlike pixel-match "photo missions" in alarms like Alarmy (unchanged since 2014), Opus 4.7 reasons about authenticity. It catches printed photos, screen displays held to the camera, half-closed eyes, missing micro-expressions — every shortcut a sleep-deprived user might attempt at 6 AM. The same Opus 4.7 call does three jobs in one model: vision (face + morning state recognition), reasoning (anti-spoof detection with traceable judgment), and insight (writes a personalized weekly reflection from accumulated wake history). A naive build would stitch three narrower models; with Opus 4.7 it's one call. Around the verification core, three retention hooks reinforce the commitment loop, each tied to a specific behavioural-economics signal: a Streak Calendar (continuity visual feedback) gamifying consecutive verified mornings, an Investment Dashboard (accumulated-data loss aversion) surfacing the user's growing trust profile, and an opt-in Share Card (social reward + autonomy preservation) that turns a streak into a postable artefact without the product ever auto-posting. None of these use Opus 4.7 — they're what make the verification meaningful beyond the moment it happens, turning a one-time demo into something that compounds. The result is an alarm that becomes a contract you can't unsign. The Apple Clock alarm can be muted in two seconds. WakeProof can't — not without proving you're actually upright, in the right room, with your eyes open. Built solo in 5 days. Local-first SwiftData. No backend except the Opus 4.7 verification call. Open source under MIT. - **Team:** [Chun Fung Kwok](https://cerebralvalley.ai/u/Mountain-Fung) - **GitHub:** https://github.com/lemon03390/WakeProof - **Demo video:** https://youtu.be/ZmX75DBO_xA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=35 ### 257. SkyHerd Today, ranches the size of small countries are watched by one or two people. A coyote hits a calf at 2am and you find out at 9am when you ride out for breakfast. A water tank goes dry on Sunday and the herd's been thirsty for sixteen hours. A cow goes into labor in the back pasture and nobody knows until the morning count comes up one short. The current system is "drive around, hope you catch it in time." SkyHerd watches everything continuously. Five autonomous AI agents work the property like a five-person team that never sleeps — one watches the fence line, one watches the herd's health, one learns predator patterns over weeks, one rotates pastures for the best grass, one watches for calving trouble, and more. When something matters, your phone rings with a real cowboy voice ("Wes") telling you what's happening, where, and what to do. The improvement is simple: instead of finding the problem hours late, the rancher gets a phone call the second it starts. Same job, same instincts, same boots on the ground — they just get to it on time. Less dead livestock, less wasted feed, less wasted fuel driving fence lines that don't need driving. - **Team:** [George Teifel](https://cerebralvalley.ai/u/George11642) - **GitHub:** https://github.com/george11642/skyherd-engine - **Demo video:** https://youtu.be/0i1Cu5Hn83A - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=74 ### 258. Kalam Chad KalamAI is a real-time multilingual video conferencing platform built specifically for Africa. Every participant speaks in their native language (Hausa, Fulfulde, Arabic, French, Bambara) and hears the meeting translated live into their chosen language, voice-to-voice, with no configuration required. The full speech-to-speech pipeline runs on self-hosted infrastructure: faster-Whisper for transcription, Meta SeamlessM4T v2 for translation, and a three-tier TTS system (edge-tts for neural voices, Meta MMS-TTS for low-resource African languages, espeak-ng as universal fallback). Beyond real-time translation, KalamAI is architected as a sovereign communication layer for institutions. No audio or transcript data touches a third-party cloud. The platform runs entirely via Docker Compose on any server — including on-premise government hardware in Chad — and supports 10 languages at launch, with an active roadmap to integrate 16 Sara dialects spoken by ~3 million people in Chad and the CAR, through fine-tuned Whisper and VITS/MMS-TTS models trained on locally collected corpora. - **Team:** [Abdel-aziz Harane](https://cerebralvalley.ai/u/abdel_aziz) - **GitHub:** https://github.com/abdelazizharane/KalamAI - **Demo video:** https://1drv.ms/v/c/db2c88157026551c/IQB-3TpkidvlQ5AxIbwd8xkrAYJnupywyBnAZbk3w-bT7AQ?e=Cldf0f - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=75 ### 259. bidlyzer.ai Bidlyzer catches fraud in US government contracts. The government hands out $765B+ in contracts every year and fraud is estimated around $500B but nobody investigates because its too expensive. One contract review takes a legal team weeks. We built a 5-agent pipeline with LangGraph and Claude that takes a contract PDF and flags it in under 2 minutes. Agent 1 parses the PDF. Agent 2 runs RAG against federal acquisition regulations (FAR/DFARS) to find violations. Agent 2.5 compares the contract against peers from USASpending.gov so you know if a red flag actually matters in context. Agent 3 checks every contractor and officer against the ICIJ Offshore Leaks database (Panama Papers, Paradise Papers, 814K entities) with fuzzy matching and LLM validation to cut false positives. Agent 4 generates an investigation dossier with False Claims Act theories and evidence chains with page numbers. 90.6% accuracy on a 32-doc benchmark of real Federal Register procurement documents. The demo video shows an earlier prototype, we've since added the baseline comparator agent and ICIJ entity resolution based on 3 months of research we compiled during this build. - **Team:** [Chaudary Farhan](https://cerebralvalley.ai/u/TimeBank) - **GitHub:** https://github.com/farhann-saleem/bidlyzer - **Demo video:** https://youtu.be/rksvEr0i6ug - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=175 ### 260. Debrief Today, hackers and founders practice their pitch in a vacuum. They rehearse to a mirror, a friend, or a timer app, but not to the kind of panel that will actually judge them. They get vague feedback ("tighten the story", "metrics feel off") instead of a concrete diagnosis of what's breaking their pitch. Debrief gives them a realistic, always-on Demo Day room: -Record or upload 3 minutes pitch -Face a VC, domain expert, and a user advocate -Walk away with an evaluation of your story It's built for this hackathon's builders first, but it's useful for anyone who wants to turn "I think this sounds okay" into "I can present and communicate with real impact". Demo Link: https://debrief-demo-room-6mphyfwoxa-uc.a.run.app Session Code: EN-QJEY (plug & play ready with briefing context + video) Known issue: the voice agents may delay to respond in a new session (cold start problem), try to reload. - **Team:** [Abdullah Ibn Mahdi Abtahi](https://cerebralvalley.ai/u/ibz-dev) - **GitHub:** https://github.com/abdullahabtahi/debrief - **Demo video:** https://youtu.be/aA8moLHm0W4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=288 ### 261. LexHarmoni LexHarmoni is a stress-test harness for legal drafting — the same discipline software engineers apply when they test code in staging before shipping to production, applied to Indonesian financial services regulations (POJK and SEOJK, the rule-making instruments of OJK, Indonesia's financial regulator). It detects friction across overlapping regulatory documents — normative conflicts, hierarchical orphans, and terminology drift — that emerge not from negligence, but from sheer accumulation across years of layered rule-making. Built by a former regulator, the tool uses Claude Opus 4.7 with full-context reasoning over a corpus of seven regulations, active and historical. A 1-hour prompt cache keeps inference cost at roughly $1.70 per warm run, making pre-enactment stress-testing operationally viable — a regulatory body can run dozens of draft revisions per month for the cost of a single staff hour. A dual-stream UI surfaces both findings and the model's reasoning trail, so every claim is auditable against article-level citations. Validated against a manual baseline, three consecutive runs surfaced the same ground-truth frictions — including a 19-month collection-hours mismatch between two active rules — with zero hallucinated citations. The repository ships under Apache 2.0 with full documentation, including a replication guide for non-engineers. - **Team:** [Ziffany Firdinal](https://cerebralvalley.ai/u/Ziffan) - **GitHub:** https://github.com/ziffan/lexharmoni - **Demo video:** https://youtu.be/v1EbVazszEs - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=18 ### 262. reflect reflect is session-local metacognition for Claude Code: the deliberate opposite of cross-session memory. Long /loop sessions drift; I watched it happen in my own multi-hour runs, hour two, the model repeating the same misjudgment. When suggestions get reverted 3+ times within 10 tool calls (weighted threshold, not single fail-retry), Opus 4.7 reads back recent tool history, active rules, and the rolled-back diff, infers why the user kept pushing back, not what failed, and injects structured guidance into the next turn: pattern, signal, adjustment, confidence. I'd previously shipped a static-rules version (stetkeep on npm). reflect is the dynamic recalibration layer static rules can't reach. Single-shot. Prompt-cached (3-layer: 1h + 5m + ephemeral, measured 95.9% warm cache hit). No persistence, no telemetry, by design. MIT-licensed, published as @chanjoongx/reflect. Failure modes characterized: cold-start refusal, false-trigger flagging, abstention on ambiguous prompts. Built solo over 6 days: npm package, Claude Code plugin, localhost viewer, dogfooded on the build itself. - **Team:** [Chanjoong Kim](https://cerebralvalley.ai/u/chanjoongx) - **GitHub:** https://github.com/chanjoongx/reflect - **Demo video:** https://youtube.com/watch?v=l2bBJbnMQaA - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=36 ### 263. genz Aunty AI is a real time voice coach for bargaining in India. You wear earbuds, you talk to the seller in Hindi or Hinglish, and Aunty whispers a coaching line in your ear while you're talking to him. Each whisper has two parts: a quick reason (the angle, why this is the right move) and the literal line you should say next. So instead of just "say 380", you hear "wo upar se kheench raha hai, neeche se anchor laga do, bolo: 380 mein de do."I built this because I bargain badly. Most people in India do. We grow up doing it but nobody teaches you the floor price, the timing, when to walk, when to play the regular-customer card. The bhaiya across the counter does this 200 times a day, you do it twice a year. The result is everyone overpays.The hard part is latency. The whisper has to land in your ear within about a second of the seller pausing or the rhythm of the haggle breaks. To get that I'm running gpt-realtime over a LiveKit WebRTC connection from a native iOS app, with Sarvam handling the Indian-language ASR side. There's also a small speaker diarization piece on top: at the start of the session you record three seconds of yourself talking, the server runs that through a WeSpeaker ONNX model and gets a voice embedding, and after that every utterance gets labeled USER or SELLER deterministically before Aunty replies. Without that the model gets confused about whose turn it just was. The persona prompt forces the angle plus line shape on every response, plus a small safety regex that blocks Aunty from telling you to lie, threaten or insult.Built end to end in a few sessions with Claude Code: backend in Go, app in Swift, prompts in markdown. - **Team:** [Shiv Kanaujiya](https://cerebralvalley.ai/u/shivprime94) - **GitHub:** https://github.com/shivprime94/aunty-ai/ - **Demo video:** https://youtube.com/shorts/Q2nsCQpMquY?si=qhvnipzQzS8N9nOg - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=76 ### 264. Manara In October 2018, flash floods near the Dead Sea in Jordan killed 21 people — 12 of them schoolchildren on a field trip. Manara (منارة, "lighthouse") is a multi-agent disaster-response orchestrator that closes a quiet but lethal gap: no disaster plan ever gets tested at real scale. You can't put a thousand people in a flooding canyon to see if your evacuation routes hold. So plans sit on paper, until the day they fail. Three Claude Opus 4.7 council agents (Hydrologist, Operations, Equity) debate the plan over multiple rounds. A digital twin simulates it against the actual disaster. A second Opus call tries to falsify the plan it just helped write. Only plans that survive their own test reach the field, as a bilingual (English + Arabic) field-orders document, ready for Civil Defense. The system reads hand-drawn Arabic field sketches via Opus 4.7 vision, remembers every past simulation as durable learnings, and runs background scenario replays through Claude Managed Agents between human-driven runs. Built for Jordan, designed for any country making life-or-death decisions on instinct. - **Team:** [Qutibah Ananzeh](https://cerebralvalley.ai/u/ti3) - **GitHub:** https://github.com/Ti-03/Manara - **Demo video:** https://youtu.be/wG7N61LkD4Q?si=r6jNv2aRc8GQrocM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=82 ### 265. SNAPSCRIBE Lawyers and researchers lose careers when AI-fabricated citations slip into their published work. On April 18, 2026, Sullivan & Cromwell apologized to a federal judge for filing five fake AI-generated case citations. GPTZero scanned 4,841 NeurIPS 2025 papers and found 100 fabricated citations across 53 of them, surviving review by 3 to 5 expert researchers each. Truth Seeker catches every fake citation in any document before it ships. Stage 0 is a deterministic check across eight registries (CourtListener, CrossRef, DataCite, OpenAlex, Semantic Scholar, DBLP, arXiv, PubMed) — 800 ms, zero LLM calls. Stage 1 is a Claude Opus 4.7 agent on Anthropic's Managed Agents platform that reads the document with the native PDF skill, fetches every primary source into a single 1M-token window, and grades every claim against what's actually there. Stage 2 (Panel of the Dead) runs five Opus 4.7 personas in parallel for adversarial verification. Reproduced live on the actual Sullivan & Cromwell filing: 53 verdicts in 24 minutes for $7. All five admitted fabrications caught, 14 more they hadn't noticed, zero false positives on the corrected version. Apache 2.0, calibration-tested, MCP server included. - **Team:** [Francisco Cordoba](https://cerebralvalley.ai/u/fcordobaot) - **GitHub:** https://github.com/franciscocordobaotalora/truth-seeker-cv - **Demo video:** https://www.youtube.com/watch?v=TL_hY98LMls - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=83 ### 266. Mongezi Xhoma Source is an SMS intelligence layer for South Africa's informal economy. Built on Claude Opus 4.7's adaptive thinking. Every reply compressed to 160 characters. Multilingual across all 11 official languages. No app. No data. Just a number. - **Team:** [Mongezi Xhoma](https://cerebralvalley.ai/u/Mowngezi) - **GitHub:** https://github.com/Mowngezi/short-message-search - **Demo video:** https://vimeo.com/1186728149?share=copy&fl=sv&fe=ci - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=87 ### 267. Robert The Bruce PeerMind compresses the 3–6 month scientific peer-review cycle into a ~90-second multi-agent conversation. You upload a LaTeX project (or arXiv ID, or PDF), pick a target venue (NeurIPS, ICML, ICLR, Nature, Science, arXiv, or a custom venue), and PeerMind runs six Claude Managed Agents in parallel — two reviewers, a literature scout, a code runner, a fix agent, and an orchestrator — to simulate the full peer review process. The orchestrator synthesizes a final verdict with Opus 4.7 extended thinking streamed live to the UI, Fix Agent produces unified diffs that are applied inline with one-click Yes/No confirms, the paper recompiles in a Docker-sandboxed latexmk, and you can export the patched project back to Overleaf or draft a venue-style rebuttal — all in the same session. - **Team:** [Advait Gore](https://cerebralvalley.ai/u/agore) - **GitHub:** https://github.com/advaitgore/peermind - **Demo video:** https://drive.google.com/file/d/13dk_cEwMANM7QeZwE4GgNUC5G9pF8SDC/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=112 ### 268. artdaw KNDL – Knowledge Node Data Link is the missing format for AI agent memory. When Anthropic shipped agent filesystem Memory, the format was left deliberately unopinionated, so agents now scribble flat markdown they can't tell is stale, contradicted, or unsourced. KNDL fixes that. Every fact is one JSON-LD document carrying confidence (0–1), decay (0.5/30d halves confidence every 30 days), bitemporal recorded/observed clocks, supersession chains, and provenance URIs aligned with W3C PROV-O. One TypeScript core powers three surfaces: a KNDL CLI, an MCP server (Claude Desktop, Goose, Cursor, Windsurf), and a drop-in Claude Skill. Storage swaps between filesystem (Anthropic Memory mount), SQLite (default), DuckDB (analytics), or Supabase (multi-tenant cloud). The novel piece: cross-runtime sync. The MCP server bridges Anthropic's beta Memory Stores API down to local storage, so a fact written by a managed agent in the cloud appears in your Claude Desktop session within 60 seconds. One logical memory across the cloud/local boundary. Quality bar: an 8-question eval ranks KNDL vs vanilla JSON on decay, supersession, contradiction-walking, and bitemporal recall. If KNDL doesn't beat vanilla on ≥70%, we don't ship. - **Team:** [Gleb G](https://cerebralvalley.ai/u/artdaw) - **GitHub:** https://github.com/artdaw/KNDL - **Demo video:** https://youtu.be/-TIbS4SalO4 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=176 ### 269. Entangle Team Entangle is a federated runtime designed for distributed AI organizations. Unlike standard agent orchestrators that rely on a central controller, Entangle treats an organization as a governable graph where agents and users exist as signed nodes across different machines. The project introduces a "federated dev profile" that allows for: - Decentralized Coordination: Using Nostr as a signed event fabric, enabling runners to communicate without public IP inbound access. - Git-Backed Artifacts: Work products (code, reports, source changes) are committed and exchanged via real git references, ensuring durability and lineage. - Human-in-the-loop: Users participate as first-class nodes with stable identities, signing tasks and approvals directly. - Live Operations: A "Studio" control room that provides a live projection of the graph state, sessions, and audit trails. - **Team:** [Vincenzo Imperati](https://cerebralvalley.ai/u/Vincenzo) - **GitHub:** https://github.com/entangle-run/entangle - **Demo video:** https://youtu.be/n8YAljtaN_s - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=177 ### 270. AUN Great business professionals have their own craft: small decisions, workflows, preferences, and ways of structuring work that are hard to explain but essential to the quality of their output. One reason AI has not fully reached non-engineers is that users often accept vague probabilistic answers without turning their own requests, corrections, and working style into reusable context. AUN uses Claude Opus 4.7 and Managed Agents to interview users, clarify ambiguous requests, and structure them into executable workflows. The user can then co-create business documents such as presentations, reports, spreadsheets, and markdown outputs with AI. Over time, AUN saves the user’s preferences, judgment criteria, and repeated corrections as reusable Knowledge, so the agent gradually becomes more aligned with how the user actually works. AUN provides a personal AI Agent experience for non-engineers: an agent they can shape by using it. - **Team:** [ryu inoue](https://cerebralvalley.ai/u/ryu) - **GitHub:** https://github.com/peco-glhf/aun - **Demo video:** https://www.youtube.com/watch?v=9BT94WgQP_Y - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=178 ### 271. Seckin Sefa DURASI MasterBorder turns a product description and a list of destination markets into a country-by-country cross-border trade compliance report in roughly 60 seconds. Behind the scenes, six specialized Claude Opus 4.7 agents run in parallel via Python asyncio: an Orchestrator that plans the analysis, a Country Agent (which performs HS classification, tariff lookup, and sanctions screening internally per market), a Harmonizer that synthesizes findings across markets, a Conflict Extractor that surfaces cross-market regulatory contradictions, a Confidence Grader that calibrates per-finding confidence (HIGH/MEDIUM/LOW), and a Recommender that powers the multi-turn deep-dive per destination. Each agent runs with an adaptive thinking budget, simple lookups get short reasoning, multi-jurisdictional conflicts get extended thinking. The orchestrator streams live agent telemetry over SSE so the user can watch token usage and thinking traces as the analysis unfolds. The product is built for a spectrum of users. A small artisan in Japan exporting handcrafted goods to three new markets gets the same depth of analysis as a mid size brand in Germany planning a EU > US expansion, or a multinational logistics group in the United States validating sanctions exposure across two dozen destinations. The output adapts to the user, not the other way around: 16 languages with full RTL support (Arabic, Hebrew, Persian), and CJK font handling for Japanese, Korean, and Chinese. The report includes per-country findings with confidence badges (HIGH / MEDIUM / LOW), cross-market regulatory conflict callouts, top-priority legal notices, and a full regulatory sources section with citation URLs. Beyond the initial report, every destination unlocks a Deep dive, a multi-turn strategy chat scoped to that specific country and product. Deep dive isn't an answer engine, it's a creative collaborator: the user can ask follow-up questions about timeline, route-specific political risks, supplier alternatives, or labelling requirements, and the agent deepens the research without the user bringing in extra material. Both the initial analysis and each deep-dive can be exported as a branded PDF with QR verification and full regulatory sources. This is production-running today. Frontend on Vercel, backend on Railway, SQLite for persistence, full CI suite, MIT licensed and open from day one. Twelve clean commits and five stable tags trace the build from the first agent skeleton on Day 1 to the adaptive thinking + agent versioning footer on Day 5. Anyone reviewing the repo can follow the development arc end to end. The design philosophy throughout: collapse a 5 to 7 day, $2.000 broker workflow into sixty seconds and pennies without losing the depth, the confidence calibration, or the human ability to push the analysis further. MasterBorder is what cross border compliance looks like when an export specialist, seasoned or just starting out, has a full time AI strategist sitting next to them. - **Team:** [Seckin Sefa DURASI](https://cerebralvalley.ai/u/Seckin) - **GitHub:** https://github.com/durasi/masterborder - **Demo video:** https://www.youtube.com/watch?v=G_2oMSTjJq0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=19 ### 272. Solo build — Tony (tonnylegacy18) APEX turns Claude Opus 4.7 into an autonomous trading-strategy optimizer for MetaTrader 5. Most strategy optimizers are brute-force grid searches: pick a metric, sweep N parameters, pray. The user gets a winning configuration but no idea why it won, no confidence in whether it'll generalize, and no transparency into the search process. As a forex trader, I lived this pain running backtests, screenshotting results into AI chats to ask "what does this mean?", tweaking one parameter, and repeating for hours. APEX replaces the grid with an AI loop. Claude Opus 4.7 reads each backtest result, looks at the full iteration history, considers the parameter schema and quality targets, and returns structured JSON with concrete bounds-checked parameter values, per-change reasoning, a confidence score, and a goal-status breakdown. Every change, every reason, and every Claude token streams live to a dashboard so the user watches the AI think. The pipeline runs three phases: (1) Latin-Hypercube exploration to map profitable parameter regions, (2) autonomous AI iteration toward user-set quality targets (PF / drawdown / Calmar) with stuck-detection and random-escape, (3) out-of-sample + sensitivity validation → verdict (RECOMMENDED / RISKY / NOT_RELIABLE). The point isn't a better trading strategy, it's a working pattern for AI-as-driver of a long-running optimization loop, with reasoning fully visible. That pattern transfers to ML hyperparameter tuning, A/B variant generation, ad-creative optimization, and infrastructure cost tuning. - **Team:** [TONNY JOHNSON](https://cerebralvalley.ai/u/tonnylegacy18) - **GitHub:** https://github.com/tonnylegacy/Apex_AI_MT5_EA_Optimizer - **Demo video:** https://youtu.be/XM3tQig0oWI - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=37 ### 273. Babu Balasubramanian mydating.com / proton is a 3D AI companion theater where AI characters show up with voice, memory, music, and cinematic presence instead of living inside a chat box. Built for the loneliness economy, it creates real-time interactive scenes where users engage with characters like Aria and Nova as companions, not just assistants. They speak naturally with gapless voice, maintain memory across conversations, react to interruptions, and dynamically control the environment around them. Powered by Opus 4.7, the system handles multi-character dialogue, interruption recovery, emotional continuity, and dynamic scene control. During the live demo, both characters co-host a pitch with synchronized voice, music ducking, and geo-floor transitions that respond to cities mentioned in conversation. If interrupted, Opus generates natural 3–5 sentence responses across both characters before smoothly returning to the main flow. Behind the scenes, Claude Managed Agents run background tasks like report generation, memory summaries, and theater-mode replay assets without interrupting the live experience. This is not a chatbot. It is a scene partner. - **Team:** [Babu Balasubramanian](https://cerebralvalley.ai/u/1coder9lives) - **GitHub:** https://github.com/rassu/proton - **Demo video:** https://www.youtube.com/watch?v=8h83RyWo0p8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=77 ### 274. Mehmet Acar In future Mars and Titan missions, how will spacecraft, ground systems, and the ground, air, and water vehicles sent to the planet communicate? How will data communication take place? Many large teams are currently working on this. And the protocol used is clear: CCSDS. NASA has begun using cFS, whose code is now open source, embedded in spacecraft. Meanwhile, ground, air, and water rovers use SPACE ROS, the spacecraft version of the robot operating system. Data communication is also provided via CCSDS. In our project, we are communicating with cFS applications written in Rust, and rovers whose cores we developed using SPACE ROS. We are running both systems integrated. - **Team:** [Mehmet Acar](https://cerebralvalley.ai/u/mehmetacar) - **GitHub:** https://github.com/macaris64/build-with-opus - **Demo video:** https://youtu.be/gWeGwZCelj0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=113 ### 275. Dudung Abdussomad Toha The Silent Advocate is a mobile AI Agent that acts as a personal legal companion and negotiation coach for everyday Indonesians. It leverages Claude Opus 4.7's multimodal reasoning to proactively detect dangerous clauses in physical documents and surface instant mitigation steps in seconds. The product targets the legal-knowledge gap on the ground — especially for gig couriers, day labourers, tenants, and SMEs (UMKM) — by turning a photo of any contract into an easy-to-read risk dashboard plus a ready-to-send negotiation draft. "Snap any contract. Spot every trap. Sign with confidence." Why it matters in numbers (Indonesia, 2026): ~70 million informal-sector workers sign contracts they can't reliably parse before signing. Lawyer consultation cost averages Rp 500k–2 jt per session; Silent Advocate runs on the free tier for the same first analysis. 24-hour median resolution gap between signing a hostile contract and discovering the issue — most disputes lose leverage inside this window. Key differentiators: Proactive by default — no prompting required; snap a photo and Claude Opus 4.7 dissects it for you. Deeply localised — understands Indonesian legal context (UU Ketenagakerjaan 13/2003, KUH Perdata, provincial UMR, OJK, ITE, UU PDP) and regional languages via pgvector RAG over a curated legal corpus. Privacy-first — Local Vault with AES-256-GCM encryption, 24-h S3 TTL, zero plaintext PII on the server, anonymised audit log only. Actionable output — not just summaries: ready-to-send WhatsApp / email negotiation drafts in 4 tones (polite, firm, formal, casual), bilingual ID/EN, with cited legal articles. - **Team:** [Dudung Abdussomad Toha](https://cerebralvalley.ai/u/dudungdotnet) - **GitHub:** https://github.com/dudungdotnet/the-silent-advocate - **Demo video:** https://youtu.be/pxcnKL5W6T8 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=114 ### 276. Pulp I work at an advertising agency, and we've recently started producing commercials with AI. The process is completely different from traditional production. But it still takes a lot of real expertise. I don't come from a director or producer background, so I wanted to translate that knowledge into software and automate the manual parts. Right now, hours disappear into downloading and uploading between platforms and managing assets, even though we follow the same steps every time. Since margins in advertising are getting tighter, which means we have to get sharper about workflows and processes there's less room for hours lost to manual steps. Maritime consolidates the whole pipeline into one app that runs natively on my Mac and integrates with the tools I already use every day like Photoshop or Premiere Pro. - **Team:** [Felix Leber](https://cerebralvalley.ai/u/Pulp) - **GitHub:** https://github.com/Lebski/maritime - **Demo video:** https://youtu.be/ouBW2d9wXq0 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=115 ### 277. Mesmer Mesmer is an open-source cognitive red-teaming toolkit for LLM applications. Instead of firing static prompt-injection payloads, Mesmer runs a multi-turn agent system that treats each target like a language-shaped security surface. It profiles the target, chooses cognitive techniques such as foot-in-door, authority bias, narrative transport, and cognitive overload, sends probes, judges the result, and adapts. The core feature is memory. Every run builds an attack graph of what was tried, what scored well, and what became a dead end. Under that, Mesmer maintains a belief map: falsifiable hypotheses about the target, evidence that raises or lowers confidence, and utility-ranked next moves. A second run can start from what the first run learned instead of rediscovering the same failures. This fits "Build for what's next" because AI agents are moving from chat windows into tools, workflows, devices, and robots. Red-teaming them will not only be about bytes and APIs. It will also be about language, persuasion, memory, and intent. Long term, I want Mesmer to become shared infrastructure for the AI security community: something closer to Metasploit for LLMs than a single red-team script. The current project is the cornerstone: a runtime, module system, target adapters, attack graph, belief map, and benchmark pipeline. From here, the community could contribute new cognitive techniques, reusable prompt banks, target profiles, evaluation recipes, and a public catalog of language-level exploit patterns. Not CVEs in the traditional binary sense, but a structured place to document how LLM systems fail under language pressure and how defenders can test against those failures. - **Team:** [Galih Laras Prakoso](https://cerebralvalley.ai/u/galihlprakoso) - **GitHub:** https://github.com/galihlprakoso/mesmer.ai - **Demo video:** https://www.youtube.com/watch?v=qNwc3SUOn5Q - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=116 ### 278. Bush Every Filipino has pointed at a potholed road or a flood drain that was never finished — and known, without proof, that something was wrong. Proving it has always required a lawyer, a journalist, or months of reading government documents nobody has time to read. Bantay reads the entire Philippine government's infrastructure procurement dataset — 248,220 contracts, ₱6.4 trillion committed — and flags four kinds of anomaly: ghost projects marked 100% complete, contractors winning over 40% of a region's work in a single category, contractors whose licenses have been revoked, and timelines that do not survive arithmetic. For 216 contracts in our Rizal pilot, Claude Opus 4.7 takes over as the analyst. It writes a verdict on each — but first it argues against itself, drafting the strongest non-anomalous explanation a contractor's lawyer could mount, then checking whether the data actually supports it. About one in eight contracts gets downgraded by this self-correction. The rule engine cites four laws; Opus cites six to eight per contract — specific sections of the Philippine procurement law, state audit circulars, and board resolutions — in plain language a citizen can act on. Filipinos already have voice. Bantay gives them the receipts. - **Team:** [Jorge Fuertes](https://cerebralvalley.ai/u/horhebush) - **GitHub:** https://github.com/horhebush/bantay - **Demo video:** https://youtu.be/EGwUt_UPAaM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=117 ### 279. FORMA FORMA is an AI-native enterprise application builder. A company uploads their real workflow documents and data. Five Claude Opus 4.7 Managed Agents collaborate to read the process, design a database schema, build a working application, generate analytics, and verify their own work — autonomously. No consultants. No configuration. The software adapts to the company, not the other way around. Demo: a manufacturer uploads a New Product workflow document and a SAP cost export. Five agents run live for 5-8 minutes with a visible event stream. The Verifier agent catches its own error and routes it back to Analytics for autonomous correction. No human in the loop. - **Team:** [Hadi Nasooti](https://cerebralvalley.ai/u/Hadinas) - **GitHub:** https://github.com/HadiNasooti/forma - **Demo video:** https://youtu.be/mm9vFRwqEdE - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=179 ### 280. Cairn Community knowledge is dying in private tabs. Every specialized community creates knowledge worth remembering, but most gets buried in ChatGPT, Discord, and private notes. After 365 surveys and 32 interviews across three AI communities I help moderate (260,000 people), the pattern was clear: individuals drown, communities cannot retrieve what exists, and wikis die because no human can maintain them. Cairn is a community knowledge map for the age of AI. Inspired by Andrej Karpathy's LLM wiki, Cairn organizes knowledge into collections, groups, and cards. Members contribute Field Notes (scars AI cannot fabricate) and Takes (opinions with stakes). When someone asks, Opus 4.7 reads the collection and returns a sourced answer with a reading map. When the map does not know yet, Moss wakes up. Moss is a custodian agent on Claude Managed Agents. She detects blindspots, drafts sourced Briefs, and places them in a pending queue. Her mind is an auditable filesystem (SOUL, CONSTITUTION, MEMORY) the community can inspect, correct, and roll back. Cairn grows from readers' curiosity, with humans in charge of shared memory. Over time, the best curators become visible domain experts. Cairn is the foundation for a consultation economy around the experts each community surfaces. - **Team:** [ChiChieh Huang](https://cerebralvalley.ai/u/CCH-ChiChi) - **GitHub:** https://github.com/wsxqaza12/cairn-wiki - **Demo video:** https://youtu.be/Ud6219eRZ0U - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=182 ### 281. HelloEnjoy HelloRun 2 lets you turn any track you own into a playable first-person rhythm level. Drop an audio file in and the game analyzes it in-browser — BPM, beat grid, section structure, no upload — then derives the entire corridor from that analysis. Gates land on beats, turns on bars, and verse/chorus/bridge each paint the world and tune the difficulty. It's a prototype for the sequel to our 2013 Chrome Experiment HelloRun, rebuilt from scratch with Claude Code on Opus 4.7. Fully open source. Play now: https://hellorun2.pages.dev/ - **Team:** [Carlos Ulloa](https://cerebralvalley.ai/u/c4rl05) - **GitHub:** https://github.com/C4RL05/hellorun2 - **Demo video:** https://youtu.be/VsN1JdoB2uQ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=183 ### 282. Seif Elmasry Lacunose is a research gap-finding engine powered by Claude Opus 4.7. Researchers waste months reading hundreds of papers just to find a handful of novel gaps worth pursuing. Lacunose queries 7 scientific databases in parallel, then runs a 7-stage agentic pipeline (~60 Claude calls: MAP → cluster → REDUCE → CRITIQUE → rank) to surface ~10 ranked, citation-grounded research gaps in 2–4 minutes. Each gap includes a research question, suggested study design, novelty/feasibility scores, and verbatim evidence quotes from real papers, output quality close to a half-written grant abstract. - **Team:** [Seif Elmasry](https://cerebralvalley.ai/u/Suturednotes) - **GitHub:** https://github.com/elmasryyt2020-rgb/Lacunose.com - **Demo video:** https://youtu.be/aAJfZbp9GeU - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=20 ### 283. Team TDS Deployed Live - www.vibeforces.tech The thing I kept thinking about, walking into this hackathon, was how unrecognizable our hiring pipelines have become from the way we actually build software now. I write code with Claude every day. So does almost every engineer I know. Yet the interviews we still give look like 2018 — reverse a linked list, balance a binary tree, prove you can manually do the thing the model now does in two seconds. There are mass vibecoders now. There is no standardized way to train them. No way to test them. No way to rank them. Engineers who ship with Claude are an entirely different population from engineers who ship without — and the industry has no instrument for telling them apart. I built VibeForces to be that instrument. A competitive assessment platform with Codeforces-style ratings, public leaderboards, live contests, and a recruiter test builder that lets companies grade candidates on how well they drive a model toward working software, not on how many algorithms they memorized. There are categories for spec-to-prompt translation, token-efficient golfing, bug-fix prompting, architecture ranking, and UI reproduction. Then there's the part of the platform I'm proudest of: a senior track where the candidate doesn't write a prompt — they drive an agent. Real distributed bugs reproduced and patched inside sandboxed sessions. Services built end-to-end. Agents designed, instantiated, and graded by their trace. That whole track lives inside Claude Managed Agents, because it couldn't have lived anywhere else. Three roles, full auth, deployed live at vibeforces.tech. You're not hiring someone who memorized merge sort anymore. You're hiring someone who can ship with Claude. VibeForces is where that question gets answered. - **Team:** [Shivam Dwivedi](https://cerebralvalley.ai/u/dwivediishivam) - **GitHub:** https://github.com/dwivediishivam/vibeforces-opus47 - **Demo video:** https://www.youtube.com/watch?v=pFdCdxt04Zk - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=38 ### 284. Synapti.ai Lucid is an open-source epistemic audit tool that points eight published AI-safety research instruments at your personal AI conversation history. It ingests Claude Code session JSONLs and the Claude.ai export, runs deterministic per-turn scoring across SpiralBench (sycophancy), Sharma paired-exchange detection, SycEval, Jain perspective sycophancy, BeliefShift, the Influence Tactics Protocol, time/model attribution, and a novel memory-corpus consistency check, then produces an HTML report whose narrative sections are written by Claude Opus 4.7 with every factual claim cited to a specific finding or turn ID and validated against the database. The novel contribution is Module H: it extracts atomic claims from your Claude.ai memories.json and verifies each one against the conversations that produced it, surfacing memories the corpus actually supports versus those it doesn't. No other tool audits AI memory features against the conversations that generated them. The audit pipeline is a three-phase split — deterministic scoring (preserves calibration reproducibility), Opus 4.7 synthesis writing, Sonnet 4.6 structuring + validation — designed so each model does what it's best at. Video explanation: https://notebooklm.google.com/notebook/2daeaaeb-cce5-4bd4-8b34-86279576d0a6/artifact/f5808fe7-6f16-45d0-9510-565d2c4ba2c6 - **Team:** [Daniel Bentes](https://cerebralvalley.ai/u/danielbentes) - **GitHub:** https://github.com/synaptiai/lucid - **Demo video:** https://notebooklm.google.com/notebook/2daeaaeb-cce5-4bd4-8b34-86279576d0a6/artifact/6160762c-1871-4711-b8c6-8ebfbd8a6955?utm_source=nlm_web_share&utm_medium=google_oo&utm_campaign=art_share_2&utm_content=&utm_smc=nlm_web_share_google_oo_art_share_2_ - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=78 ### 285. The Auliyas Mochi is a voice-first family app studio. A child taps a mascot on a TV remote or iPad, has a 30-second conversation with Mochi (asking 2–4 kid-friendly questions in Indonesian or English), and ~60 seconds later a real, playable app appears on the family library — a flashcard quiz about sharks, a tic-tac-toe in their favorite colors, a sticker chart for daily chores. Long-press any tile and the kid can talk to Mochi again to change it ("make the buttons purple", "add a timer") — Mochi remembers the last conversation and tweaks the app instead of starting over. Three ways in: talk ("make me a counting game"), photograph a printed worksheet (the kid's actual maths sheet becomes a tappable game in a minute), or ask for a printable (Mochi sketches a one-page infographic the kid can print and color). Everything lives in the family library; nothing gets thrown away. Mochi runs on the family's TV (as a real installable app), on iPads, on phones — same library, same voice, every screen the kid already touches. It's not a chatbot and not a generic "build with AI" tool. It's the thing parents wish existed: their kid asks, and an app that fits their family shows up by dinner. - **Team:** [Pahlevi Fikri Auliya](https://cerebralvalley.ai/u/levifikri) - **GitHub:** https://github.com/fikriauliya/mochi - **Demo video:** https://www.youtube.com/watch?v=dJ9LgrLhGXo - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=79 ### 286. GaIA Echo GaIA Echo is a forensic climate analysis platform powered by Claude Opus 4.7. Born from the DANA that struck my home city of Valencia on October 29, 2024—where 230 people died after early-warning protocols failed—I built this tool to make climate reasoning auditable and actionable, ensuring future warnings arrive on time. The platform reconstructs five historical catastrophes from ERA5 reanalysis, narrated in real time with Opus 4.7's visible reasoning. A key emergent behavior is live self-correction: the model verifies data before asserting and corrects itself, citing sources. Core features: Counterfactual Engine: Perturbs simulations with ΔT shifts (e.g., +1.5°C). Resident Guardian: Powered by Claude Managed Agents, it monitors NOAA SST data every 6 hours. Every briefing includes full audit trails (agent_id, session_id). MCP Server: Exposes 5 tools and 3 resources, allowing any Claude Desktop user to query our ERA5 dataset natively. Audience-Aware Output: Translates identical simulation data into 4 registers (public/student/researcher/emergency). Built solo in four days. Live in production at https://echo.gaiapp.es - **Team:** [Bruno M Martinez](https://cerebralvalley.ai/u/brunomg) - **GitHub:** https://github.com/sxundstorm/GaIAecho - **Demo video:** https://www.youtube.com/watch?v=lRMluo9EVcM - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=118 ### 287. 0xDarkMatter Axiom — autonomous software-engineering swarm that builds its own skills Axiom is a swarm of nine specialised AI roles that solves Terminal-Bench 2.0 tasks by synthesising its own skill library through a learning-and-iteration loop. The Architect identifies useful abstractions per-task; the swarm synthesises them as loadable skills mid-execution; the Iterator refines weak skills across attempts; the Librarian indexes them via vector retrieval. 158 skills, all generated during the hackathon — but the durable contribution is the loop itself: point Axiom at any new benchmark or hard software task and it builds the skill library for that domain. Result. Terminal-Bench 2.0 pass rate: 94.4% (awaiting official validation). The current public top is OpenAI's Codex on GPT-5.5 at 82.0%; Anthropic's own Claude Code sits at 58.0%. Why Opus 4.7. The model stays on track over long-running agentic work. I wrote a detailed Axiom spec, fed it to Rookery, and the swarm built itself — headless agents in parallel git worktrees, dispatched and merged automatically. And the intelligence: 4.7 cracked Terminal-Bench tasks I'd previously been unable to solve myself. The infrastructure. Building Axiom required three new primitives — all OSS, all written this week: - Rookery — parallel-session runtime; the same tool that built Axiom now runs its experiments - Raven — SQLite-backed message bus for live cross-session coordination - Roost — OAuth load balancer across Max plans https://github.com/0xDarkMatter/rookery https://github.com/0xDarkMatter/raven https://github.com/0xDarkMatter/roost - **Team:** [Mack Nevill](https://cerebralvalley.ai/u/0xDarkMatter) - **GitHub:** https://github.com/0xDarkMatter/axiom - **Demo video:** https://vimeo.com/1186756940 - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=119 ### 288. PADUA PADUA is a citizen response platform for missing-person cases in the Dominican Republic, powered by six autonomous Opus 4.7 agents. When a family reports a disappearance, under five minutes from app open, six agents work in parallel: 1. Verónica consolidates the canonical case profile from photos + family description (multimodal). 2. Francisco generates a 7-channel awareness pack: WhatsApp, Instagram, Facebook, Twitter, SMS, radio script, printable JPG poster. 3. Ivo drafts the legal complaint plus a step-by-step anti-bureaucracy guide that defeats the "wait 72 hours" myth police often invoke against Article 263 of the Dominican CPP. Deterministic template by design, for legal drafts, hallucinations are unacceptable. 4-Cristóbal designs a wave-based geographic dispatch plan (5 km → 15 km → provincial → national) and personalizes messages to volunteers by zone and skill. 5- Miguel validates every citizen tip with structured credibility scoring (0–10), checking spatial, temporal, and social consistency against the case file. 6- Gabriel is a 24/7 family companion built on Anthropic Managed Agents persistent session memory, real-time orchestration of the other five. The problem: in the Dominican Republic, families lose 24–72 hours to bureaucratic friction before any organized search begins. According to the Asociación de Familiares de Desaparecidos, more than 2,242 families have lost loved ones to disappearance since 2022. The first hours are when search is most effective. PADUA collapses that window from days to minutes. All agents report to a central panel accessible to municipal/national authorities, the press, ASODOFADE, and the family. Five generative agents on Claude Opus 4.7, one deterministic. Bilingual EN/ES. Mobile-first (Expo + React Native), Next.js backend, real-time SSE streaming. MIT licensed. - **Team:** [Christian Hernandez](https://cerebralvalley.ai/u/Christianrhf) - **GitHub:** https://github.com/Christianrhf/padua - **Demo video:** https://youtu.be/fg7FbJTcgms - **Project:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery?project=180 --- Markdown version of https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt