# Google I/O Hackathon: Project Gallery - **Event:** [Google I/O Hackathon](https://cerebralvalley.ai/e/google-io-hackathon) - **When:** Sat, May 23 at 9:00 AM – 10:00 PM (PDT) - **Where:** Shack15, San Francisco, CA - **Hosts:** [Cerebral Valley](https://cerebralvalley.ai/u/cv), [Google DeepMind](https://cerebralvalley.ai/u/deepmind) - **Projects:** 152 (6 placed) - **Page:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery ## Projects ### 1. CODESTREAK RoadSOS is an offline-first AI emergency response platform powered by Gemma 4 that transforms smartphones into intelligent emergency nodes capable of coordinating rescue operations even when internet connectivity and traditional infrastructure fail. The project was built to address a critical real-world problem: most emergency systems depend heavily on cloud connectivity, cellular towers, and centralized infrastructure, but real emergencies often happen in remote highways, disaster zones, rural areas, and low-connectivity environments where those systems fail first. RoadSOS combines local Gemma 4 inference, BLE mesh networking, offline emergency coordination, multimodal emergency understanding, responder intelligence, and decentralized communication into a resilient emergency-response ecosystem. Core capabilities include: • Offline SOS activation • AI-powered emergency triage • BLE mesh communication between nearby devices • Multimodal crash scene understanding • Nearby SOS community alerts • AI-assisted responder dashboards • Medical ID emergency access • Safe Walk monitoring • Transparent emergency activity logs Gemma 4 enables RoadSOS to perform emergency reasoning directly on-device, allowing: • Offline AI inference • Low-latency emergency processing • Privacy-preserving workflows • Edge multimodal reasoning • Reliable operation during infrastructure failure RoadSOS was intentionally designed for: • Disaster zones • Rural communities • Remote highways • Flood-hit regions • Low-connectivity environments • Global resilience scenarios Our long-term vision is to evolve RoadSOS into a decentralized emergency infrastructure layer integrating responders, smart vehicles, wearables, disaster communication systems, and future smart-city emergency coordination networks. RoadSOS is not just an app. It is AI emergency infrastructure built for the moments where human life becomes most vulnerable. - **Team:** [Nitish R G](https://cerebralvalley.ai/u/nitishrg) - **GitHub:** https://github.com/NITISH-R-G/RoadSOS - **Demo video:** https://youtu.be/TmKCmk0Eh_I?si=IUOEDAKKI-fZlrl5 , https://youtu.be/Xwzwmf7U8Nc - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=1 ### 2. Execution Assurance Bureau AI generated repositories increasingly optimize for plausibility rather than behavioral correctness. A repository can boot successfully, expose polished APIs, and still contain incomplete or non functional runtime behavior underneath. RepoProbe is an execution assurance system for AI generated repositories. Instead of statically analyzing code, it executes repositories, probes runtime surfaces, fuzzes endpoints, verifies behavioral consistency, and compares observed runtime evidence against functionality claimed in repository READMEs. The pipeline is intentionally split into two layers. Deterministic systems handle repository fingerprinting, execution planning, runtime probing, contradiction detection, and behavioral verification. Gemini 3.5 Flash is then used to synthesize runtime evidence and reason over contradictions without generating evidence itself. The investigation system escalates suspicious runtime behavior into deeper probing workflows. For example, if an authentication endpoint returns invariant responses across malformed JWTs, empty credentials, and randomized payloads, RepoProbe flags the implementation as behaviorally inconsistent with its claimed authentication guarantees. The terminal interface streams the investigation live as execution unfolds: runtime boot events, route discovery, contradiction escalation, fuzzing, trust score degradation, and final runtime assessment. The project was built as infrastructure for a problem that is becoming increasingly common with AI generated software: code that looks production ready while remaining behaviorally incomplete under execution. - **Team:** [Shivam Gawali](https://cerebralvalley.ai/u/shiv669) - **GitHub:** https://github.com/shiv669/repoprobe - **Demo video:** https://youtu.be/mmrrb-pUJjU?si=ibLyhYnO8C0PoLNO - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=2 ### 3. GemiSupport Autonomous AI agents that automatically escalates to humans based on sentiment and complexity - **Team:** [Henry Kendall](https://cerebralvalley.ai/u/hskendall), [Heidi Wu](https://cerebralvalley.ai/u/258258258) - **GitHub:** https://github.com/hskendall/nano/upload/main - **Demo video:** https://youtube.com/shorts/T63C8FNcuYQ?feature=share - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=3 ### 4. daylite INTERLOCK is the agentic orchestration layer for deepfake forensics at the moment of authorization — for any high-stakes voice or video call. When a deepfake CEO impersonates a real executive on Google Meet to authorize a wire transfer — the exact scenario Arup faced in Hong Kong, January 2024, losing $25.6 million in fifteen transfers — INTERLOCK orchestrates a Council of specialist signals, gates on consensus, and publishes a verdict event the bank's risk system acts on. The Council is six Gemini 3.5 Flash sub-agents — Frame Forensics, Voice-Print Cross-Match, Reverse Provenance (Search-grounded), Counter-Strategy, Regulatory Precedent (Search-grounded), and Injection Guard — fanning out in parallel from an Antigravity Managed Agent orchestrator. Frame Forensics and Voice-Print are designed as adapters that wrap specialist detectors via Antigravity sandbox tool calls — Modulate Velma (1.1% EER · HF Speech Deepfake Arena #1), Resemble DETECT-3B Omni (97.4% accuracy), Pindrop Pulse (secures 8 of 10 largest US banks). We are the diagnostic laboratory; they are the microscopes. A verdict aggregator gates on 3-of-6 consensus; no single sub-agent can block or override. The same architecture defends wire confirmations, hiring interviews, claim assessments, and KYC re-verification calls — voice or video. Containment runs as a real `interactions.create({base_agent: "antigravity-preview-05-2026"})` sandbox interaction. The sandbox executes Python that signals the bank's wire-freeze workflow, locks the CEO account through the bank's existing risk pipeline, and drafts an SEC 8-K Item 1.05 disclosure Search-grounded against EDGAR precedent (Microsoft 2024-01-17 Midnight Blizzard). INTERLOCK never debits — the bank's existing infrastructure acts on the verdict, same deployment model as CrowdStrike or Pindrop. Every irreversible action runs in a sandboxed env_id with a surfaced interaction_id — auditable, replayable, never autonomous. Architectural orientation: the Vector Institute's May 15 2026 "Generalization Illusion" report — SynthForensics benchmark shows 29.19% mean AUC drop for SOTA detectors against novel diffusion generators. Standalone media forensics is fragile; orchestrated consensus above multiple diverse specialists is the architectural answer. Six sub-agents in parallel; verdict aggregator + comms layer sequential after consensus — eight total Gemini 3.5 Flash calls per detection. Built solo in 13 hours. - **Team:** [Dmitrii Karataev](https://cerebralvalley.ai/u/kwit) - **GitHub:** https://github.com/kwit75/interlock - **Demo video:** https://www.youtube.com/watch?v=R_RwgO8CWMs - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=4 ### 5. Color My Voice Color My Voice is an engaging mobile web app that translates human speech into beautiful, frameable sound-art posters based on the historical pitch-color associations of classical composer Nikolai Rimsky-Korsakov, a reknowned synesthete. While roughly 4% of the world experiences synesthesia (blended senses), researchers lack large, diverse datasets to study it. Our app solves this by turning clinical voice collection into a fun, zero-friction creative experience. With explicit consent, it securely crowdsources anonymized, numerical voice summaries into a local research database—giving scientists the global data scale they need while keeping participant data completely private. - **Team:** [Grace Esteban](https://cerebralvalley.ai/u/gracesteban) - **GitHub:** https://github.com/mgesteban/color-my-voice - **Demo video:** https://youtu.be/iiXBZS97MAE - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=5 ### 6. AgentCanvas AgentCanvas is the first visual orchestration canvas for Gemini 3.5 Flash sub-agent pipelines. Most multi-agent systems today are black boxes - you run code, agents execute, you see a final output with no visibility into what happened in between. AgentCanvas changes that. Drag and drop Gemini agents onto a canvas, connect them into pipelines, define their roles and tools, and watch every tool call execute in real time with a live animated trace. Every agent handoff, every tool call, every output is visible as it happens. Built entirely during the hackathon. Three demo pipelines included: Research (Scout → Analyst → Writer), Productivity (Planner → Scheduler → Communicator), and Engineering (Researcher → Engineer → Reviewer). Fully customizable add agents, define roles, draw connections, run any pipeline. Tech stack: React frontend with live SSE streaming, FastAPI backend, Gemini 3.5 Flash for all agent execution. - **Team:** [Preeti Parihar](https://cerebralvalley.ai/u/itspparihar) - **GitHub:** https://github.com/preeti13parihar/agentcanvas - **Demo video:** https://drive.google.com/file/d/1lVv_6AHcWWv3UKFoBJoAgd-ppVmLxyrr/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=6 ### 7. Aegis Aegis is an autonomous purple-team engine designed to evaluate, exploit, and self-harden AI agents in a closed execution loop. Built from scratch with zero framework bloat, Aegis deploys parallel adversarial sub-agents to target a helper bot with mutated context-manipulation and tool-abuse exploits (OWASP LLM Top 10). When a vulnerability is exposed, a defensive orchestration layer dynamically analyzes the breach telemetry, synthesizes targeted runtime prompt-boundary patches, and updates the agent's context in real-time. The entire process streams into a custom interactive 3D telemetry dashboard (dashboard.html) showing the live transition from systemic breach to a hardened 3/3 blocked security posture. - **Team:** [Greg d](https://cerebralvalley.ai/u/gRRReat) - **GitHub:** https://github.com/gadaugherty/Aegis - **Demo video:** https://www.loom.com/share/d7bc86fcc85f41e295fca588c5753db0 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=7 ### 8. Voca Voca is a voice-first agent that interviews people in their own language to fill out government forms, validates every answer in real time, and reads the completed form back to them in their language before they submit it themselves. - **Team:** [Rahul Thennarasu](https://cerebralvalley.ai/u/rahulthennarasu), [Siddharth Balaji](https://cerebralvalley.ai/u/siddharth-balaji) - **GitHub:** https://github.com/sbalaji09/voca - **Demo video:** https://youtu.be/_wk1NvYQG30 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=8 ### 9. Omnidesk Omnidesk is an AI production desk for creator music videos. A creator provides a prompt, lyrics, and optional creator-owned assets, and Omnidesk turns them into a rights-safe music-video plan with two connected Veo scenes and one continuous Lyria soundtrack. Gemini 3 Flash creates the scene plan, Veo prompts, safety notes, and music structure. A four-agent Gemini Managed Agent desk reviews the plan for IP safety, video prompt quality, creative direction, and music continuity. The creator can optionally apply those agent recommendations back into the plan, ignore them, or continue straight to generation. Omnidesk generates Veo clips, produces a separate Lyria music bed, combines the final MP4, stores outputs in Vercel Blob, and publishes watchable results to a community Discover feed. - **Team:** [Spencer Yang](https://cerebralvalley.ai/u/spenceryang) - **GitHub:** https://github.com/spenceryang/omnidesk - **Demo video:** https://youtu.be/uiF7RhHPh_M - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=9 ### 10. Autonomous Infrastructure Auditor 1 in 4 Americans lives with a disability, yet most cities have no systematic way to audit pedestrian infrastructure for ADA compliance at scale. Manual inspections are slow, expensive, and cover only a fraction of the street network. The Autonomous Infrastructure Auditor solves this by deploying a fleet of AI specialist agents across any urban district to autonomously detect accessibility violations such as steep ramps, missing tactile indicators, insufficient crosswalk timing. These are things that prevent wheelchair users, visually impaired pedestrians, and ambulatory individuals from navigating the city safely. A District Orchestrator dispatches three specialist agent types concurrently, each walking real geodetic terrain and applying persona-specific ADA thresholds to measured physics. What makes the system distinctive is how it grounds every finding in multiple modalities simultaneously: real terrain elevation data drives the slope physics, live Street View imagery is analyzed by Gemini's vision capability to visually confirm hazards at the exact coordinate, and the agents' regulatory reasoning is anchored to specific ADA code sections rather than generalized descriptions. When a location fails multiple criteria at once, two specialist agents negotiate a unified fix through bidirectional context passing, producing a single remediation that resolves compound hazards together. The result gives city planners actionable, location-specific violations at a scale and speed no human audit team could match. - **Placement:** Finalist - **Team:** [Kushaan Naskar](https://cerebralvalley.ai/u/Kushaan) - **GitHub:** https://github.com/Kushaan-N/navAssist - **Demo video:** https://www.youtube.com/watch?v=CvKXWNurbSM - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=10 ### 11. Delphi Delphi · synthetic populations as a computational primitive. You ask any question — Will the Fed cut rates in Q3? Pretest this tagline. Stress-test this decision — and a swarm of Gemini 3.5 Flash sub-agents, each role-playing a different American persona generated from real US Census demographic axes and grounded in live web, reasons in parallel. In ~60 seconds you get back: a probabilistic forecast with confidence interval, the strongest reasons for and against drawn from agents' own reasoning, the demographic axes where groups diverged most, and a striking outlier quote — all synthesized into a Wall Street Journal–style summary by a final Gemini call. Live-validated on stage at N=500 with a +66.2 percentage point shock response (13.3% → 79.5% when CPI surprised). Validation across 27 automated tests, an adversarial persona-stability harness (4.92/5 in-character, 0% drift to centrist mean), and a cross-model comparison that showed Gemini 3.5 Flash holds 100% per-agent success where Gemini 2.5 Flash collapses to 12.5% on identical prompts. Until Gemini 3.5 Flash made hundreds of parallel grounded reasoning agents economical at conversational latency, this category did not exist. Forecasting is the wedge — the same primitive powers marketing pretests, policy war-gaming, synthetic juries, and behavioral pre-mortems. - **Team:** [Sankaranarayanan Subbayya](https://cerebralvalley.ai/u/Sankar) - **GitHub:** https://github.com/SankarSubbayya/delphi - **Demo video:** https://youtu.be/FObF2xpgBAw - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=11 ### 12. SkillStudio A verified registry + CLI for Gemini agent skills. Authors publish from GitHub; vskill runs an automated security/quality scan powered by Gemini Flash 3.5, assigns a trust tier (T1–T4), and lets users install pinned versions with one command. Adds version diffing and drift detection between local and upstream — so you always know if a skill is stale or unsafe. - **Team:** [Anton Abyzov](https://cerebralvalley.ai/u/EasyChamp) - **GitHub:** https://github.com/anton-abyzov/vskill - **Demo video:** https://www.youtube.com/watch?v=yEg46Ybh4Yk - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=12 ### 13. Squat Sensei Squat Sensei is an iOS squat counter that uses your iPhone camera and on-device pose detection to count reps automatically. An energetic AI coach delivers motivational voice lines powered by Gemini 3.1 Flash TTS. The app was built with Google Antigravity. Problem: Most fitness apps make you tap to count reps or need extra hardware. Squat Sensei keeps it simple — open the app, squat, and get coached in real time with expressive AI voice. - **Team:** [Shohei Yukawa](https://cerebralvalley.ai/u/Shohei) - **GitHub:** https://github.com/yukawashouhei/Squat-Sensei - **Demo video:** https://www.youtube.com/shorts/hsUBTbPaj7w - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=13 ### 14. AgentLab When a multi-agent pipeline produces a confident wrong answer — which step broke, and how do you know? AgentLab is an evaluation system for multi-step agent pipelines. It watches production traces, detects quality drift, and does the part nobody else automates: it localizes the regression to the specific step that caused it. From the real failing traces it drafts targeted evaluation cases; a human approves them; the fleet is re-scored against the enlarged suite; the degraded step is rerouted to a healthy agent - **Team:** [Apoorv Shrivastava](https://cerebralvalley.ai/u/Bot-007) - **GitHub:** https://github.com/apoorv-shrivastav/agentlab-mesh-router - **Demo video:** https://drive.google.com/file/d/1AsSD_0KjbnIqE5P1pSbGFa0hR4xlWsfE/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=14 ### 15. dshenoy_EvoAI EvoAI is a proactive ambient spatial memory companion that watches your environment silently, builds a living memory of every object and its state, and recalls anything on demand via voice — no tags, no manual logging, no friction. The problem it solves: We constantly misplace things — keys, wallet, charger, food, medication. Current solutions require active tagging (AirTags need to be attached to every object) or constant cloud surveillance (privacy invasive, always-on cost). EvoAI watches passively with a local camera, only fires an AI call when something actually changes (motion-gated), and answers naturally: "Where did I leave my pouch?" → "It's on your desk, to the right of your monitor, in the study." "Did I close my laptop lid?" → re-inspects the stored frame visually and confirms what it actually sees. - **Team:** [Datthesh Shenoy](https://cerebralvalley.ai/u/datthesh_Shenoy) - **GitHub:** https://github.com/dattheshshenoy/Google_IO_Hackathon/ - **Demo video:** https://drive.google.com/drive/folders/1vuFajn9A5AdRkjqQgKJwxS83n4I-IdO4?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=15 ### 16. kopard 2026: AI has solved production. A founder can build a real product in a weekend — the code is no longer the bottleneck. Distribution is. Finding the people who actually need what you built is harder than building it. LaunchAgent is the answer. LaunchAgent reverses the marketing pipeline. Instead of writing content and chasing audiences, it finds conversations where developers are already asking for your product, then drafts a platform-specific reply for each — landing your message inside an active discussion instead of broadcasting blindly. You type one sentence about your product. 12 parallel Gemini 3.5 Flash agents fan out: 5 run in isolated Antigravity sandboxes via the Managed Agents API and execute Python + curl against Reddit, Hacker News, GitHub, Dev.to, and Stack Overflow; an aggregator and a dedicated relevance filter clean the demand map; a strategy agent picks targets; then 6 reply agents draft replies that target the highest-leverage comment in each thread and quote the actual commenter's words. Two broadcast agents draft X and LinkedIn posts. A Nano Banana 2 social card and a Veo 3 fast 8-second product video round out the assets. The Launchpad gives you one-click "Open thread + copy reply" — opens the actual Reddit / HN / GitHub comment permalink in - **Team:** [Dr Dhanraj Chavan](https://cerebralvalley.ai/u/dermdc) - **GitHub:** https://github.com/drdhanrajchavan-dotcom/find-where-demand-is - **Demo video:** https://drive.google.com/file/d/1EjjbuxqaJGsxENuDEpYnpp94e9lxqCx_ - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=16 ### 17. world-cup-io A feature-rich, browser-based tile exploration and simulation game with integrated AI and managed agent capabilities and a built-in level editor. Deployed link for tryout: https://mochiyaki.github.io/app4/ - **Team:** [sam sam](https://cerebralvalley.ai/u/gguf) - **GitHub:** https://github.com/mochiyaki/world-cup-io - **Demo video:** https://github.com/mochiyaki/world-cup-io/blob/main/demo.gif - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=17 ### 18. ZoneLyft.ai Zonelyft.ai is an autonomous environmental compliance platform that uses a multi-agent AI pipeline to perform full NEPA (National Environmental Policy Act) audits and generating EIS (Environmental Impact Statement) in minutes - a process that today takes atleast 2.5 years and costs millions. Every real estate and infrastructure project in the United States that involves federal land, federal funding, or significant environmental impact is legally required to file an Environmental Impact Statement before construction can begin. A single EIS takes 4–7 years on average, costs $2–10 million, and is done almost entirely by hand - consultants manually reading regulations, cross-referencing satellite imagery, pulling permit histories, and writing a formal legal document that can run over 1,000 pages. Zonelyft.ai solves this by automating the entire workflow. A developer submits a parcel location and project description. Five AI agents - powered by Gemini 3.5 Flash on Google ADK - immediately get to work: one analyzes 8 years of satellite imagery from Google Earth Engine to establish the environmental baseline, one searches 2M+ indexed federal regulations to identify what applies and what is violated, one retrieves 30 years of permit history and EPA violation records, one cross-references all of it against the developer's own claims to detect contradictions, and one drafts the formal NEPA document in statutory language. The result is a compliance determination, a conflict matrix, a regulatory findings table, and a downloadable draft EIS - in under few minutes. The problem it solves is not just speed. It is honesty. Today, developers self-report their environmental impact and regulators catch errors months or years later. Zonelyft.ai grounds every finding in satellite truth that cannot be disputed, regulatory text that is publicly searchable, and an immutable audit trail that ties every conclusion to its source. - **Team:** [Vaibhav Vemula](https://cerebralvalley.ai/u/vaibhaaaavvv), [Bhoomika Nanjaraja](https://cerebralvalley.ai/u/Bhn) - **GitHub:** https://github.com/vaibhav-vemula/ZoneLyftAI - **Demo video:** https://www.loom.com/share/2e91c33af32f44959feeae309b92f980 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=18 ### 19. Mojo 3 Do you have anything wanna do that have been hover on my mind but never get it done? You’re not lazy, just need a few more breakdowns. Mojo cat is here to guide you to clarify the path, breakdown to the minimal action unit, use gamify scheme to provide instant positive feedback that easily keep you push the loop. We use the psychological CBT method, shield the noise in your brain to make you focus at present, leverage the instant feedback loop to create dopamine to keep you move forward. - **Team:** [Elaina YAO](https://cerebralvalley.ai/u/wanting_yao), [Haiting Diao](https://cerebralvalley.ai/u/dev_hdiao), [Luona Guo](https://cerebralvalley.ai/u/Luona) - **GitHub:** https://github.com/wanting4/mojo-1.3 - **Demo video:** https://youtu.be/ShG5LTv7WIY - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=19 ### 20. ClinAgent ClinAgent is a multi-agent clinical decision-support system built on the Google Antigravity SDK and Gemini 3.5 Flash. From a patient's EHR, wearable trends, a clinician's voice dictation, and uploaded medical images or documents, a team of specialized AI agents works in parallel to draft and verify clinical documentation — a SOAP note, prior-authorization request, and discharge summary. Gemini 3.5 Flash's low latency and native multi-modality let every agent run concurrently and reason over images and PDFs directly. Layered safety checks surface missing data and allergy conflicts without ever silently blocking the clinician. - **Team:** [Yash Barapatre](https://cerebralvalley.ai/u/trailblazerYB), [senay yakut](https://cerebralvalley.ai/u/senay) - **GitHub:** https://github.com/Google-I-O-CV-Hackathon-2026/clinagent - **Demo video:** https://drive.google.com/file/d/1k77uIXPsmDfQSdXToUra1XKTO9dIbMmL/view?usp=drive_link - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=20 ### 21. Gemini Lifeguard Rescue Gemini Lifeguard Rescue is an AI-powered water safety system that helps lifeguards detect and respond to emergencies faster. The demo focuses on beach rescue, using three drones to patrol the coast, inspect possible hazards, and deliver flotation support when needed. The same idea could also be used at pools with fixed cameras and video monitoring as well as flotation support. The system works by having Gemini act as the mission commander. One drone patrols the area, another checks possible emergencies with live video, and a rescue drone can be sent if flotation support is needed. Gemini analyzes the situation, helps classify the risk, and recommends the next action for lifeguards. AI and robotics are the future of safety. This project is a step toward using intelligent machines to help prevent fatalities during swimming, beach, and other water activities. - **Placement:** Finalist - **Team:** [Tommy Purcell](https://cerebralvalley.ai/u/tommypurcell) - **GitHub:** https://github.com/tommypurcell/Google-I-O-Hackathon - **Demo video:** https://youtu.be/ox_PaFAm7SE - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=21 ### 22. team-swag Build a library of reusable skills from code you've already written. Feed them to coding agents to ship faster. Like a producer reusing loops they've made before, mainstacks lets you extract patterns from past projects and drop them into new ones. Your agents get context on how you build things, so they stop guessing and start building the way you would. - **Team:** [collin hargreaves](https://cerebralvalley.ai/u/cjhhhh) - **GitHub:** https://github.com/cjhargreaves/mainstacks - **Demo video:** https://streamable.com/1v73kl - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=22 ### 23. GTerm - Small Model Big Bench How high can Gemini 3.5 Flash climb on Terminal-Bench? GTerm is an attempt to answer this question. We believe we can beat SOTA! Early results are promising. How it works: GTerm is running a Meta-Harness-inspired self-improving loop on Flash. The original example in the Meta-Harness paper reached 79.2% with Claude Code + Opus 4.7 improving over ten continuous days of harness evolution. We’re on Day 1 with Flash and started at 37% and are climbing. Smaller model. New loop. Big bench. - **Team:** [Dominic Domoah](https://cerebralvalley.ai/u/dominicMolt), [Aditya Advani](https://cerebralvalley.ai/u/ninjaa) - **GitHub:** https://github.com/Phantastic-AI/gterm-bench-hack - **Demo video:** https://youtu.be/gOtqXY01FS8 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=23 ### 24. Hedge in a box Full ticker analysis doing the same work a hedge fund team would do leveraging agents with a human in the loop to create an investment strategy - **Team:** [Andre Miranda](https://cerebralvalley.ai/u/And) - **GitHub:** https://github.com/andremir/GoogleIOAndreHackC - **Demo video:** https://drive.google.com/file/d/18j7EDnzKAHpPyvkt5QRNVf-pZ7CLUlqa/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=24 ### 25. Heat Threshold Heat Threshold answers one question: should I go outside, and when? You type (or speak) a location, an activity, and a time. A primary Gemini 3.5 Flash agent dispatches a graph of ten sub-agents — three uploaded as persistent Managed Agent definitions, one weather cascade (NWS → Open- Meteo), and six deterministic sub-agents for routing, sun-shade math, refuge scheduling, and Street View. It returns a go / delay / alternate verdict grounded in Stull (2011) wet-bulb math and USMC 6200.1E flag thresholds. Same plan, three surfaces: a 2D bento dashboard, a playable WebXR HUD over Google Photorealistic 3D Tiles, and a Gemini Live voice mode that calls our Managed Agents pipeline as a tool — bridging the two newest Gemini APIs through function calling. Every run is traced by PlatAtlas, recorded to Vercel Blob by McpTape, and replayable via McpReplay (?replay=) — the demo safety net. Built today, 10:30 AM – 5:00 PM PT. MIT. Environmental scheduling tool, not medical advice. - **Team:** [craig merry](https://cerebralvalley.ai/u/craigmerry) - **GitHub:** https://github.com/HeatThreshold/HeatThreshold - **Demo video:** https://youtu.be/9qWiDOWgPXs - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=25 ### 26. Oneline Oneline: software for one. Describe the exact tool you need in one line, and get it running on your phone in under a minute, built for the single person asking instead of a market. A planner trims the request, three candidate tools build in parallel inside Gemini Managed Agents sandboxes, a three-layer judge scores each (functionality via Playwright inside the sandbox, UX clarity, design coherence), a deterministic selector picks the winner, and it deploys to Cloud Run with a QR code. Each losing candidate's mistakes become lessons that improve the next build. Not an app builder like Lovable or V0 which ship to a market; Oneline builds a disposable tool for the one person who asked, used, and throws it away. - **Team:** [Chanjoong Kim](https://cerebralvalley.ai/u/chanjoongx) - **GitHub:** https://github.com/chanjoongx/oneline - **Demo video:** https://www.youtube.com/watch?v=Lm403MpBo_Y - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=26 ### 27. Forger Forge is a product-first agent system that helps people discover, evaluate, and build product opportunities. It supports both connected products, like an existing GitHub repo, and new-product ideas. The system ingests project context, manual ideas, repo signals, public research, and user preferences, then ranks opportunities, runs critique and Bull/Bear review, recommends a next action, and lets the user approve a direction for prototype or MVP generation through the dashboard. It then autonomously builds it for them and opens a PR so the demo can actually be tested. - **Team:** [Nadav Avital](https://cerebralvalley.ai/u/nadav), [Ron Kibel](https://cerebralvalley.ai/u/rkibel) - **GitHub:** https://github.com/nadavital/Forge - **Demo video:** https://drive.google.com/drive/folders/1QuJovKH6PyKu33YnOA0U6zYUd26IqC7k?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=27 ### 28. TinyLM Swarm Lab TinyLM Swarm Lab is a local mission-control app for running tiny language-model research experiments. The local orchestrator uses Gemini 3.5 Flash to generate experiment hypotheses, then assigns independent workers to test different tiny-LM configurations. Each worker runs real training code, produces artifacts like config.json, metrics.json, notes.md, train.log, and checkpoint.json, and reports results back to the dashboard. The project solves the manual loop of small-model experimentation: choosing hypotheses, launching runs, reading logs, comparing metrics, and deciding what to try next. TinyLM Swarm Lab makes that process visible and repeatable by showing the active worker fleet, current best model, efficiency trade-offs, artifact outputs, and checkpoint-based sampling in one UI. - **Team:** [Ethan Zemelman](https://cerebralvalley.ai/u/ethanzemelman) - **GitHub:** https://github.com/Zemelware/TinyLM-Swarm-Lab - **Demo video:** https://drive.google.com/file/d/1KfzhOo1UVWTn1SVoDUs1Lvy8dha3lOX-/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=28 ### 29. Groundstate AI Groundstate enables the productivity of a modern AI agent without leaking sensitive data to the cloud. Sometimes, you just can't upload critical information to AI on the Cloud. We leverage local models to identify tool needed and sanitise API calls, while using advanced agents capabilities online. Your secrets are safe, while enjoying the commodity of the latest Gemini Agents! Ideal for doctors, lawyers and high stakes professionals. - **Team:** [Matthieu gsa](https://cerebralvalley.ai/u/mgsa11) - **GitHub:** https://github.com/mgsa1/Groundstate - **Demo video:** https://www.youtube.com/watch?v=t7fm7IuNNj4 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=29 ### 30. Claims - Live Claims Live - A Flash Agent. Always Ahead. It is a real-time, Gemini 3.5 -powered agent that joins every claims intake call alongside the human rep. It listens, structures, and submits — turning a stressful phone call into a compliant, fully captured claim before the caller hangs up. /n #The problem /n When someone files a First Notice of Loss, they're rarely calm. A multi-vehicle collision on the freeway. A water heater that just flooded the basement at 2am. Stranded at an airport after a hurricane diverts every flight. Emotional distress is at its peak — and on the other end of the line is a claims rep who has to do four hard things at once: Capture under pressure. Chronological event descriptions, vehicle makes and models, damage segments, loss locations, contact details — all extracted from a customer who is mid-sentence, repeating themselves, and emotionally overwhelmed. Reason against policy in real time. Coverage limits, deductibles, named exclusions, riders, endorsements. The rep has to cross-reference the customer's story against a dense policy document while the customer is still telling the story. Stay compliant. Identity verification, disclosure of deductibles, explanation of timelines, mandatory empathetic acknowledgment of loss. Miss any of these and the claim is exposed. Read the room. Distress level, pacing, tone. Push too hard and the caller breaks; move too slowly and they hang up. Under this load, two things happen: details fall through the cracks (the wrong vehicle, the missing photo, the unrecorded address), and the rep stops actually listening — because they're typing. The solution Claims Live runs as an agent on every call. It transcribes in real time, extracts structured FNOL fields as the conversation unfolds, validates them against the policy, surfaces the next compliance check the rep owes the customer, and auto-fills the claim form. When everything's captured, the agent submits the claim and returns a confirmation number — all before the caller has finished saying goodbye. The rep stops typing and starts listening. The customer feels heard. The claim gets filed correctly the first time. - **Team:** [Bharat Bhavnasi](https://cerebralvalley.ai/u/bvsbharat) - **GitHub:** https://github.com/bvsbharat/Claims-Live - **Demo video:** https://youtu.be/X6p6YjwXsko - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=30 ### 31. Agentory CAO, the Chief Agentic Officer, is a full-stack company runtime where a founder can turn one messy goal into coordinated work across specialized Gemini managed agents. Instead of another chatbot, CAO models a company as nine accountable AI departments: CAO, Engineering, Design, Sales, Marketing, Operations, Support, Finance, and Legal. The system routes work, runs department agents, saves their outputs as artifacts, tracks approvals, records run replay/audit history, and connects to external tools through integrations like Composio. Our demo shows a founder goal becoming a real operating plan: CAO routes the request, a managed department agents (powered by google's new custom managed agents and Gemini models) executes the work, the output is saved to the company library, and the admin console proves the run with health, approvals, usage, replay, and audit records. - **Team:** [Shreedhar Pandey](https://cerebralvalley.ai/u/i-only-vibe-code) - **GitHub:** https://github.com/Shree-git/CAO-chief-agentic-officer - **Demo video:** https://youtu.be/kTbgBxgwT1I - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=31 ### 32. Legitfire Legitfire is a multi-agent application that answers one consumer question: is this legit? Bona AI authenticates the communication against a dozen vast numbers of data registries and threat feeds, returns a cited verdict, and then Legitfire pursues the case across every reporting jurisdiction it touches. - **Team:** [Joel Koh](https://cerebralvalley.ai/u/j1x) - **GitHub:** https://github.com/onextech/legitfire - **Demo video:** https://youtu.be/8iEUtEMFWJ4 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=32 ### 33. carte Carte is a self-generating digital business card: type a business name and it returns a micro-site that reflects that brand's identity. Palette, type, copy, real photography of the venue. Any business can now generate a polished card on demand. - **Team:** [Adam Torres Encarnacion](https://cerebralvalley.ai/u/adamrt) - **GitHub:** https://github.com/AdamPSU/google-io - **Demo video:** https://youtu.be/_0AlJZDx5DY - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=33 ### 34. Watsons Watsons are automated product analysts that replaces manual data pulls with AI-driven analysis. A PM selects from a catalog of 87 analyses across 13 categories (retention, monetization, engagement, etc.) and Watsons run them end to end: pulls the data, applies the right statistical methodology, generates charts, and surfaces a ranked feed of insights worth attention, not raw numbers. They also continuously monitor KPIs for anomalies and triggers qualitative research (competitor analysis, voice-of-customer, unmet-need discovery) when a quantitative finding warrants explanation. An Insight Engine scores, deduplicates, and ranks everything into one prioritized feed so a startup can open it Monday morning and immediately see what needs attention. A reinforcement learning loop captures PM feedback on every surfaced insight and continuously tunes the ranking weights so the system gets better at predicting what deserves attention over time. - **Team:** [Anshul Dhawan](https://cerebralvalley.ai/u/anshuldhawan) - **GitHub:** https://github.com/anshuldhawan/watsonproductassistant - **Demo video:** https://youtu.be/DrOzV2qU9yE - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=34 ### 35. Billie Gene Billie Gene is a rapid vaccine-design research acceleration platform built to reduce the early bottleneck in responding to emerging viral outbreaks such as Ebola or Hantavirus. Traditional vaccine discovery can take months before researchers have well-structured candidate designs to evaluate. Billie Gene accelerates this first stage by taking a raw pathogen sequence and running it through an agent-driven scientific workflow that produces research-stage mRNA vaccine candidates in about 10 minutes. - **Placement:** 1st Place - **Team:** [Eric He](https://cerebralvalley.ai/u/eriche2000), [Victoria Wang](https://cerebralvalley.ai/u/vikang), [Kangcheng Ji](https://cerebralvalley.ai/u/k4ji), [Weining Li](https://cerebralvalley.ai/u/LWN) - **GitHub:** https://github.com/WeiningLi/billiegene - **Demo video:** https://www.youtube.com/watch?v=dvEPovQ8qlk - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=35 ### 36. Volly LLMs are typically very poor at generating ASCII art, calling into question their ability to model the real world. This project generates asciii art uses a method inspired by RL, but not technically RL that improves a prompt automatically and chooses what it thinks is the best attempt at generating an ascii art image of a specified animal. - **Team:** [Ziad A](https://cerebralvalley.ai/u/ziad) - **GitHub:** https://github.com/ziadgit/volly - **Demo video:** https://www.youtube.com/watch?v=PCGvsMUlnzE - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=36 ### 37. AnimaOS AnimaOS reduces the workload of traditional AI chats: too much text, no unified interface, and no clear way to understand or secure what the agent is doing. Instead of giving an AI access to your personal laptop, AnimaOS gives humans access to an operating system designed for AI agents: observable, constrained, and built for transparent action. - **Team:** [Marin Lebrec](https://cerebralvalley.ai/u/Yafa) - **GitHub:** https://github.com/Hodisy/AnimaOS - **Demo video:** https://youtu.be/Yrkwu1dXkag - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=37 ### 38. Cortex Axonlotl is a self-healing infrastructure platform designed to solve the critical brittleness of multi-agent LLM systems, which currently face production failure rates of 41–86%. While traditional software relies on deterministic fault tolerance (retries/catch blocks), Axonlotl implements a synthetic nervous system that treats agent networks as biological entities capable of autonomous repair. - **Team:** [U G](https://cerebralvalley.ai/u/thegroot), [Preyash Yadav](https://cerebralvalley.ai/u/preyashyadav) - **GitHub:** https://github.com/preyashyadav/axonlotl - **Demo video:** https://youtu.be/LCQKQOdytKA - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=38 ### 39. Talon Talon is a digital engineering teammate that watches real work surfaces like GitHub, Slack, health checks, and deployment signals. Instead of waiting for a prompt, it clocks in on a schedule, collects signals, remembers recent work, understands company context, decides what needs attention, and takes safe actions autonomously. - **Team:** [Ege Uysal](https://cerebralvalley.ai/u/egeuysall) - **GitHub:** https://github.com/egeuysall/talon - **Demo video:** https://youtu.be/99rGodZ4QlA - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=39 ### 40. Autonomous Player Fleet Autonomous Android Simulator Player Fleet & Telemetry Ingest It is a premium, real-time platform designed to run autonomous gameplay testing agents on a fleet of Android devices (both physical and virtual emulators) using the Google Antigravity (AGY) SDK and Gemini 3.5 Flash, while simultaneously ingesting full diagnostic telemetry, player actions, device logcats, screenshots, and agentic actions. - **Placement:** Finalist - **Team:** [Bourke Floyd](https://cerebralvalley.ai/u/chbfiv) - **GitHub:** https://github.com/bourkefloyd/google-io-hackathon-2026 - **Demo video:** https://youtu.be/11Hqb10Qtwc - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=40 ### 41. Crossmodal Search Engine A natural-language search tool for the Waymo Motion Dataset. Using Waymo’s multimodal data—camera frames, object positions, acceleration, trajectory, and scene context—the system lets users describe the driving scenario they want to find in plain language. For example: “Find scenes where a pedestrian jaywalks.” The tool then searches across the dataset and returns the exact frames or scene segments that match the requested situation, making it easier to identify complex behavioral patterns without manually reviewing hours of autonomous driving data. - **Team:** [Rae Jin](https://cerebralvalley.ai/u/raejin) - **GitHub:** https://github.com/jin-dalrae/2605-GoogleIO-crossmodal-search - **Demo video:** http://youtube.com/not yet - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=41 ### 42. Gemini Clarity I built Clairity around the managed agents pattern. Three Gemini 3.5 Flash sub-agents and gemini-3.1-flash-live-preview (for live visual) run in the background while Gemini Live handles the voice loop. One watches the camera and narrates only when something matters. One captions and embeds moments the wearer wants to remember. One classifies every wake word utterance in 150 ms. Together they feel like one assistant. - **Team:** [Aryan Farhang](https://cerebralvalley.ai/u/darkstep13) - **GitHub:** https://github.com/AryanFP/Project-V - **Demo video:** https://drive.google.com/drive/folders/1XEIly9-Hi8YbgnThx3d9FeoFhUvSe4ic?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=42 ### 43. Self Healing Agent Loop powered by Antigrativity This project implements a CI/CD-agnostic self-healing agent loop powered by the Google Antigravity SDK. It supports: Multiple CI providers (GitHub Actions, GitLab CI, Jenkins, CircleCI, manual/local) Multiple test runner styles (pytest, jest, playwright, maven/junit, go test, generic) Config layers (config file → environment variables → CLI overrides) Execution modes (local dry-run, local apply, CI autonomous, webhook server) Safety guardrails (attempt limits, allowed path constraints, patch-size limits, rollback, audit log) - **Team:** [Shivay Lamba](https://cerebralvalley.ai/u/shivaylamba) - **GitHub:** https://github.com/shivaylamba/antigravity-sdk-self-healing-tests - **Demo video:** https://www.boomshare.ai/shared/01KSBKDPY7N4GS3MVJAS9HSYVM - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=43 ### 44. GemmaQuest: Serverless WebGPU RPG GemmaQuest is a browser RPG where Google's Gemma 4 E2B (~2.3B params, Q4) runs 100% on-device via WebGPU — no server, no API, no internet. NPC dialogue and biome lore stream live from the model while the player's network is offline. The whole project ships as a single HTML file (gemmaquest.html) that contains the spec, an in-page test runner, and the playable game. ▎ ▎ The problem it solves: most "AI games" are thin wrappers over a hosted API, which means latency, cost, privacy leakage, and zero offline play. GemmaQuest proves a frontier-class open model can author a game world live, on the player's own GPU, with zero network — by toggling DevTools → Offline before loading the realm and watching the LOCAL · GPU badge stay green while NPCs keep talking. ▎ ▎ Built on transformers.js + ONNX Runtime Web (WebGPU EP) running onnx-community/gemma-4-E2B-it-ONNX. A single persistent Web Worker sits behind an InferenceEngine adapter interface with a serialized request queue, so biome-lore and NPC-chat generations never race on the GPU. A drop-in ApiEngine implements the same interface as a fallback insurance path. - **Team:** [Sterling Cobb](https://cerebralvalley.ai/u/fourcolors) - **GitHub:** https://github.com/fourcolors/google-io-2026 - **Demo video:** https://youtu.be/j_MeM_ZkIns - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=44 ### 45. MTS MomTest Studio is a platform for founders to practice talking to customers. It provides different personas and allows founders to get feedback on their conversation approach based on "The Mom Test" by Rob Fitzpatrick. - **Team:** [Richie Tan](https://cerebralvalley.ai/u/RichieTan) - **GitHub:** https://github.com/richtan/MomTestStudio - **Demo video:** https://youtu.be/O_ziCTa9rZk - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=45 ### 46. Zero Bytes AI-Powered Football Talent Discovery for Underserved Regions - **Team:** [Keerthi Surisetty](https://cerebralvalley.ai/u/ks11) - **GitHub:** https://github.com/zoeyeballard/GMI_google_hackathon - **Demo video:** https://youtu.be/jBMvqchJ4e0 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=46 ### 47. Syst.md Syst.md is a voice-driven tool that turns a spoken system design conversation into both a structured Markdown spec and a live 3D architecture diagram. You talk to a conversational AI auditor (Gemini Live, real-time audio) that walks you through four sections in order: goals, architecture, components, and data flow. After each section is confirmed, the auditor red-teams it with one sharp critique, a failure mode, scaling cliff, or security gap, forcing you to harden the design before moving on. As sections are confirmed, DESIGN.md is written incrementally and a separate interpreter agent (Gemini structured output, Zod-validated) generates a spec that drives a Three.js scene with per-type shapes, ELK-laid-out edges, and particle animations showing data flow direction. The 3D diagram updates progressively after every section so you watch your architecture build itself in real time, with a 2D/3D view toggle and fullscreen. This solves three problems at once: design conversations are usually unrecorded, adversarial thinking is rarely applied during early design, and architecture diagrams are typically static drawings made hours after the discussion all three happen automatically as you talk. The final output is a hardened DESIGN.md plus an interactive 3D visualization of the system you described, produced in a single conversational session. - **Team:** [Brandon Tautuan](https://cerebralvalley.ai/u/BrandonTautuan), [Gerard Consuelo](https://cerebralvalley.ai/u/epicgdog), [Sanskar Thapa](https://cerebralvalley.ai/u/sanskar1016pro) - **GitHub:** https://github.com/sskarz/Google-IO-Hackathon-2026 - **Demo video:** https://youtube.com/shorts/20uG4MRK-Hk?si=ZcOFgM8oynKxvoRA - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=47 ### 48. better scroll Better Scroll is a short-form video app that adds AI-generated short form content, derived from the user's bookmarks - blogs, papers, youtube watch later; to their existing short form content. We use google managed agents and veo to stitch 30 second reels for the user to consume continuously. - **Team:** [jagrat rao](https://cerebralvalley.ai/u/jagrat7), [Rajat Bansal](https://cerebralvalley.ai/u/rajatb) - **GitHub:** https://github.com/RJBansal/better-scroll - **Demo video:** https://www.loom.com/share/648c864324cf4584a79f88a8a7ca3d6d - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=48 ### 49. Mitosis Turn the conversations a team already has into self-improving workflow agents — with humans kept in the loop. - **Team:** [Ryan McNutt](https://cerebralvalley.ai/u/ryanmcnutty) - **GitHub:** https://github.com/mcnutt1414/mitosis-synthetic-company-simulator - **Demo video:** https://www.loom.com/share/ec3c75717f2741d7bc147885f95f63ac - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=49 ### 50. Mira.ai Mira is the visualization layer for thinking. You speak any question, and four Gemini 3.5 Flash sub-agents fan out to build a studio-grade animated explanation in under 30 seconds, with follow-ups that morph the scene in real time. Writing compresses mental models into text, and rebuilding them mentally is work most people skip. Authors like Bartosz Ciechanowski and Grant Sanderson solve this with hand-crafted interactive explainers, but each one takes weeks to months. Mira automates the artisanal layer so any concept can get its own: how a phantom traffic jam emerges from nothing, how an AI generates an image from a sentence, why a bicycle actually stays upright. Speak a question, see the explanation. - **Team:** [Idan Kestenbom](https://cerebralvalley.ai/u/Macintosh) - **GitHub:** https://github.com/Macintosh1011/mira - **Demo video:** https://youtu.be/aacwSzwpDKo - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=50 ### 51. Sentinel Sentinel is an agentic privacy-routing system: sensitive data never leaves your hardware, yet the cloud still gets useful work done. A Google Managed Agent ingests a document and, before reading a single raw byte itself, delegates the un-redacted content to a Gemma model running locally on the operator's NVIDIA DGX Spark (via Ollama behind an allowlisted tunnel). The local model classifies and masks every sensitive span (PII / financial / medical); the cloud agent then reasons only over the masked output and produces a PDF report plus an audit-log.json that proves, line by line, what touched cloud vs. local. The headline guarantee — verified end-to-end — is raw_sensitive_bytes_processed_in_cloud: 0. The problem it solves: organizations want cloud-scale AI but can't legally or ethically expose raw PII to it. Sentinel makes "the cloud trusts the local model's answer without ever seeing the raw data" an enforceable, auditable boundary rather than a policy promise. - **Team:** [Rich Lira](https://cerebralvalley.ai/u/richlira) - **GitHub:** https://github.com/richlira/sentinel - **Demo video:** https://www.tella.tv/video/richs-video-5sj5 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=51 ### 52. Nudge Nudge is a local-first, physical-to-digital state tracking framework. It bridges physical environments (people, objects, interactions) with digital agents. A cheap local model handles continuous, private frame-by-frame entity and pose tracking. Only when physical state changes, Nudge calls Gemini Flash for gated semantic analysis and clustering-keeping API costs near-zero and user data fully private. It then nudges digital agents or human workflows (via Telegram, WhatsApp, etc.) to trigger real-world actions like purchases. - **Team:** [Gautham Gururajan](https://cerebralvalley.ai/u/g-g), [Vinay Kudari](https://cerebralvalley.ai/u/vinay), [Stanley Mei](https://cerebralvalley.ai/u/Stanley20001208) - **GitHub:** https://github.com/vinaykudari/googio - **Demo video:** https://youtu.be/bvJgoSCbiIs - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=52 ### 53. GemCoder GemCoder is a agentic coding tool that lets developers define, run, and improve their own agentic coding harness as code. Each repository owns its harness through version controlled files. GemCoder orchestrates between Google's Managed Agents API in the cloud and the Antigravity SDK on your laptop, same engine, two runtimes, your choice per task. These files define the agent’s instructions, reusable skills, allowed context, verification commands, approval policy, and runtime routing. GemCoder turns that harness into a repo-aware workflow for the Gemini and Google Managed Agents stack, with safe context mounting, structured task packets, patch-first outputs, local verification, run graphs, and support for local or remote execution. The TUI streams agent steps and tokens live, highlights diffs, and previews patches before applying. - **Team:** [Yahya Alhinai](https://cerebralvalley.ai/u/yhinai), [Ramis Hasanli](https://cerebralvalley.ai/u/ramizik), [Shashi Jagtap](https://cerebralvalley.ai/u/ShashiJagtap) - **GitHub:** https://github.com/ramizik/gem-coder - **Demo video:** https://youtu.be/99fsrIM4ppA?si=DIJs4iBrSUQjxJ6F - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=54 ### 54. Luillyfe Strategic Search and Sourcing AI Agent for real state asset evaluation and due diligence - **Team:** [Fermin Blanco](https://cerebralvalley.ai/u/luillyfe) - **GitHub:** https://github.com/luillyfe/due-diligence-dashboard - **Demo video:** https://youtu.be/cH7AX3eLBu4 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=57 ### 55. StackMap StackMap is an AI onboarding workspace that turns a public GitHub repo into an interactive learning experience for new engineers. Paste a repo URL, pick your role (frontend, backend, full-stack, infra, QA, or OSS), and StackMap builds an interactive architecture map with services, APIs, data layers, entry points, and risks connected by relationships you can explore. Click any node to inspect summaries, key files (linked to GitHub), and flagged gaps. You get a Week-1 Path of five learning missions (not refactor tickets): trace flows, read key files, understand risks, and plan a first contribution. Missions unlock in order; you check off success criteria and mark steps complete as you go. A Familiarity Score tracks progress across architecture, frontend, backend, data, testing, infra, and risk awareness. Mentor AI chat answers questions in context about the selected node or mission, and can use a remote sandbox to inspect real source. Story Mode generates a short narrated walkthrough of the area you're focused on, with optional audio. - **Team:** [Anmol Sharma](https://cerebralvalley.ai/u/Anmols), [Siddharth Bhat](https://cerebralvalley.ai/u/Sid42) - **GitHub:** https://github.com/thesid42/Stackmap - **Demo video:** https://drive.google.com/file/d/1CSeL3bn1efFcbXqyw9XchgzuMlIXfAtQ/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=59 ### 56. Reflex Reflex is a self-improvement loop for agent harnesses. It watches real Antigravity execution traces, sends the trace plus the harness’s current skill files to a Google Managed Agent, and runs a recursive language-model reflection loop with Gemini 3.5 Flash to figure out what went wrong. The reflector can then edit an existing skill or author a new one, turning failed runs into better operating procedures for the next run. In our demo, Antigravity follows a stale skill, hits missing scripts/files/endpoints, and Reflex automatically patches the harness skills so the next agent discovers the correct project command and validates successfully. - **Team:** [Anthony Ronning](https://cerebralvalley.ai/u/anthonyronning) - **GitHub:** https://github.com/AnthonyRonning/reflex - **Demo video:** https://youtu.be/kPH6HwUFb94 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=53 ### 57. Wake (Jack Gardner solo) Wake is a world model for organizational action. Drop one action a company could take — an acquisition, a layoff, a product change, an internal proposal — into a graph of ~200 real entities (executives, competitors, journalists, customer cohorts, platforms, regulators), where every node is its own Gemini 3.5 Flash reasoner. Then watch the consequences cascade in real time: you can read who is reasoning and why, one rationale at a time, as the press turns, communities go hostile, and a private grievance leaks when a node's public face drifts too far from what it privately thinks. It doesn't stop at one run. Wake runs the same action many times with the world perturbed, collapses the futures into outcome regimes, and surfaces the single pivotal variable that most decides which one you land in. You can take the same action and see how it plays out under different framings, and click any outcome to trace the exact causal chain back to your decision, cited event by event. The world isn't fixed either — describe a scenario in plain language and Wake assembles the real cast and builds a runnable world around it. The thesis: as AI takes over execution and humans become the judgment layer, every decision-maker needs to ask "what happens if I do X?" before doing it. Robotics is building world models for physical action; Wake builds one for organizational, textual action. It is not an oracle. It makes the consequence space legible, so you can watch your decision before you make it. Gemini 3.5 Flash is what makes hundreds of parallel reasoners cheap enough to actually run. - **Team:** [Jack Gardner](https://cerebralvalley.ai/u/jackgardner) - **GitHub:** https://github.com/Mortifera/google-io-hackathon-2026-05-23-wake - **Demo video:** https://youtu.be/Gj8Rsuz7UNY - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=55 ### 58. Foodlytics https://foodwastelytics.vercel.app/ Foodlytics is an AI-powered food waste command centre that helps operators, policy analysts, and households take action against food waste — targeting a 50% reduction against a January 2026 baseline across South Asia, East Africa, and Latin America. THE PROBLEM One-third of all food produced globally is wasted. That is 1.3 billion tonnes per year, $1 trillion in economic losses, and 8% of global greenhouse gas emissions. The data to fix this exists, but it is fragmented across government reports, academic papers, and siloed databases — inaccessible to the people who need to act on it. HOW IT WORKS Foodlytics ingests live food waste interventions from the web using Gemini 3.5 Flash with Google Search grounding, stores them in a structured database, and automatically vectorises each intervention for semantic search. An AI advisor answers questions from operators and policy teams, grounding every response in the live intervention database and FAO supply chain guidelines. Every answer includes citations traceable to a real source. When an intervention underperforms, an autonomous Gemini agent investigates. It calls five tools in parallel — satellite crop anomaly feeds, demand signals, FAO loss rate coefficients, waste event history, and a semantic knowledge base — then writes a corrective action brief with a measurable 30-day recovery target. A household waste tracker lets users log discarded food with automatic CO2 and water impact calculations. The AI advisor uses this personal data to give tailored recommendations, not generic tips. A fridge scan feature is designed but not yet connected to the frontend. Users would photograph their fridge and receive expiry risk estimates and zero-waste recipe suggestions powered by Gemini Vision. TECHNOLOGY Built on Next.js 16 and Vercel for the frontend, Supabase with pgvector for the database and vector search, and Gemini 3.5 Flash with Gemini Embedding-001 for AI generation and 1536-dimensional semantic embeddings. Foodlytics is not a dashboard that shows you the problem. It is an action engine that finds interventions, monitors them, investigates failures, and grounds every AI response in real citable data. - **Team:** [Hironmoy Paul](https://cerebralvalley.ai/u/hironmoy), [Mitil Roy](https://cerebralvalley.ai/u/Mitro), [Tarlan Isaev](https://cerebralvalley.ai/u/organicnz) - **GitHub:** https://github.com/organicnz/foodlytics - **Demo video:** https://www.youtube.com/watch?v=t1sa-_8dcZc - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=68 ### 59. Reading Copilot Reading Copilot is a hands-free AI reading assistant for physical books. Using Meta Ray-Ban or phone camera capture, it lets readers ask questions about the page in front of them, translate pointed words, and automatically build a lightweight memory of what they have read. The problem we solve is that physical books are hard to search, summarize, or ask contextual questions about without breaking reading flow. Reading Copilot turns short, event-based page captures into page-grounded answers, summaries, page updates, and a “So Far” memory trail. - **Team:** [Young Han](https://cerebralvalley.ai/u/younghan_meta) - **GitHub:** https://github.com/seyeong-han/ReadingCopilot - **Demo video:** https://youtu.be/uc1HKosogyU - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=77 ### 60. Dreamstate Dreamstate Simulator is an AI-powered agent simulation that turns dream narratives into visible emotional systems. Users input a dream, and Gemini helps extract symbolic and emotional signals from the text. These signals are mapped onto a population of simulated agents with varying sensitivity, resilience, trust, and suggestibility. As the simulation runs, users can observe how uncertainty, fear, repair, and withdrawal spread through the network, inspect individual agent reasoning, and test counterfactual interventions that stabilize the system. The project explores how subjective inner experience can be modeled as emergent collective behavior through explainable AI and interactive simulation. - **Team:** [Jeni Dhodary](https://cerebralvalley.ai/u/jenidho) - **GitHub:** https://github.com/jenidho/dreamstate - **Demo video:** https://youtu.be/0nb66zQiL5M - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=118 ### 61. Gods Eye Cloned successfully to /Users/deepanshgandhi/Documents/gods-eye. Sentinel (gods-eye) — Description A fully AI-driven global surveillance command center demo. Every layer of the pipeline is generative AI — synthetic footage, per-camera analyst agents, and a 3D Earth interface. An operator orbits a real Google Photorealistic 3D Earth with camera pins anchored to actual lat/lon. Each pin glows: 🔴 red — active crime 🟡 yellow — suspicious 🟢 green — clear Clicking a pin shows the AI-generated CCTV clip, the agent's reasoning trace, and a recommended action. - **Team:** [Deepansh Gandhi](https://cerebralvalley.ai/u/deepansh1404), [Benjamin Merchin](https://cerebralvalley.ai/u/BenjaminTop), [Deon Menezes](https://cerebralvalley.ai/u/Deonmenezes) - **GitHub:** https://github.com/deonmenezes/gods-eye - **Demo video:** https://drive.google.com/drive/folders/1IextP3-vbE8UAstnjF7TTZ6qHbNqwCNB?dmr=1&ec=wgc-drive-hero-goto - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=121 ### 62. Done with Flash Done solves the issues with managing complex family schedules with dozens of tools. Proactive agent execution keeps everyone sane and in sync. - **Team:** [Michael Bergin](https://cerebralvalley.ai/u/higharc) - **GitHub:** http://github.com/michaelbergin/done - **Demo video:** https://youtu.be/U6SVm5fcaR8 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=140 ### 63. Forge Hardware work is gated by expert knowledge: when a board won't come up you need a power engineer, a signal-integrity engineer, and a firmware engineer over your shoulder — and you rarely have all three. Forge saves companies countless productivity hours and replaces those people with agents. It watches your bench through passthrough video (supporting meta quest AND iphone), listens to your question, and convenes a guild of specialist, Gemini-powered agents that deliberate in parallel — each one autonomously pulling the relevant datasheet and board-doc facts before it speaks. When they disagree, Forge surfaces the dissent in a live panel so you see the competing hypotheses before you act, then converges on safety-gated, step-by-step instructions with every value (voltages, current limits) cited against the board's datasheet, not guessed. A snapshot escalates to a stronger vision model to read chip markings and wiring. Built on Gemini Live (voice + vision), Gemini 3 vision, Gemini function-calling/agent SMEs, and a LangGraph-based orchestrator — running live on a Meta Quest in mixed reality. - **Team:** [Bhavik Sheoran](https://cerebralvalley.ai/u/Bhavik), [Alex Hernandez](https://cerebralvalley.ai/u/alexhernandez) - **GitHub:** https://github.com/hernantech/googleiohackathon - **Demo video:** https://youtu.be/EIXd_FrP6Ok - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=148 ### 64. Gemini Display Kit Three commands ship a Gemini-powered game on Meta Ray-Ban Display. Meta's smart-glasses Web App surface just opened up. It's a 600×600 HUD driven by Neural Band swipes — powerful, but it can't hold an API key, open a camera, or call Gemini directly. Every developer trying to build something good on it has to reinvent the same stack: a bridge server, an SSE pipe, a Cloudflare tunnel, an install QR, and wrappers for every Gemini surface. Gemini Display Kit packages that stack as a CLI + SDK. npx gdk create my-game --template adventure && gdk dev boots the bridge, mints the install QR, and the game is on the lenses in 90 seconds. The framework treats Gemini Managed Agents as the creative spine, not a chat box. The repo ships real captured code_execution_call traces in-tree at examples/prompt-arena/data/traces/m_emberton.json — a managed agent that wrote Python and ran it in a remote Linux sandbox to tune a monster's battle strategy. Scrollable proof of server-side code execution, persisted state, and structured output. That's the managed-agents prize claim, committed to git. 
Two original games ship as live demos: OMNI-ODYSSEY is a generated illustrated adventure. Gemini Flash narrates each beat with memory across choices; Nano Banana paints every scene on demand; Veo opens it with a cinematic. The audience plays too — they scan a QR for /conjure, describe a monster in plain English, Nano Banana paints it, and it appears in the live on-stage story as an encounter their character must face. Pure generation; vision input never used. PulseBlade360 is an original 360° wearable rhythm game. Gemini Flash designs the level structure as JSON, Lyria scores the backing track, a managed agent reviews the level for balance. The player turns and swipes as targets enter the hit zone. Not Beat Saber; not an image analyzer. Built 100% during the hackathon (2026-05-23). Every commit timestamp is within event hours; PROVENANCE.md is the file-by-file ledger. The only non-original code is @google/genai and qrcode npm deps. Google APIs used: Gemini 3.5 Flash (narration + structured JSON), Gemini Managed Agents via the Interactions API with code_execution_call and persistent threads, Nano Banana (gemini-3.1-flash-image-preview) image generation, Lyria 3 music generation, Veo image-to-video for cinematics, Gemini TTS, and the @google/genai JavaScript SDK. - **Team:** [Sid Kandan](https://cerebralvalley.ai/u/sidk) - **GitHub:** https://github.com/sidkandan/GeminiDisplayKit - **Demo video:** https://www.youtube.com/shorts/6Gl1k9jtep4 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=56 ### 65. prompt as code Convert code, Python for now, to natural language prompts. The prompts can be used later on to generate code with similar functionality in other languages also. - **Team:** [S G](https://cerebralvalley.ai/u/whatdhack) - **GitHub:** https://github.com/whatdhack/prompt-as-code - **Demo video:** https://github.com/whatdhack/prompt-as-code/blob/main/README.md - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=70 ### 66. Mad Max MADMAX: Autonomous Market Ecosystem is a multi-agent corporate war-room simulation built for the Google I/O Hackathon 2026. Powered by the Gemini 3.5 Flash Managed Agents API, the system pits three rival AI corporations against each other in a hyper-accelerated, zero-human-latency digital wasteland. A top-level orchestrator dynamically deploys specialized sub-agents—handling R&D, pricing warfare, and corporate espionage—communicating entirely through high-velocity, structured JSON payloads. When operators trigger catastrophic macroeconomic shocks, the ecosystem instantly adapts, showcasing emergent defensive strategy, real-time multi-step reasoning, and extreme token optimization under pressure. - **Team:** [Nik Adlakha](https://cerebralvalley.ai/u/NikDWise), [Soumya Bhaiya](https://cerebralvalley.ai/u/soumyabhaiya), [Kaustubh Gharat](https://cerebralvalley.ai/u/kaustubh1504), [Harshita Jogi](https://cerebralvalley.ai/u/HarshitaJogi04) - **GitHub:** https://github.com/Kaustubh1504/leaderboard - **Demo video:** https://drive.google.com/file/d/1sb63zfaUt-4qP41SiAj5sycYPv7K487M/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=63 ### 67. MarketResearchAgents Demographic Reaction Engine predicts how the US population would react to a pitch, ad, product page, or policy — before you ship it. Paste copy or a URL. Google Gemini returns a structured analysis across five demographic dimensions (age, education, region, gender, marital status) with stance scores from -100 to +100, concerns, benefits, and verbatim quotes. Those segment reactions are mapped onto 13,354 simulated personas, rendered as a stance histogram, archetype heatmap, polarization index, and browsable persona directory. Persona dataset. Built from NVIDIA Nemotron-Personas-USA (Census-grounded synthetic personas), filtered to working-age, and enriched with personal income via k-nearest-neighbor lookup against the 2023 ACS PUMS (~3.4M US Census records) on state × occupation × education × sex. The bigger picture. Synthetic Census personas are the open-data proof of concept. The same engine plugs into first-party customer profile data at companies like Google or Meta — turning their audience graphs into a production market research product: run any new ad, feature, price, or policy past millions of agent-simulated real customers in minutes, before a single focus group. - **Team:** [Reena Agrawal](https://cerebralvalley.ai/u/reenaagra), [Amol M](https://cerebralvalley.ai/u/akaal) - **GitHub:** https://github.com/bluefalcon2357/demographic-reaction-engine - **Demo video:** https://www.youtube.com/watch?v=Pfj-31A9pj0&feature=youtu.be - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=80 ### 68. Stay Alive Offline app to prepare you for trips to places with dangerous animals with a focus on Australia. Before the trip: - Allows you to download offline maps and Gemma - Tells you which species are dangerous in the area, what they look like, what to do if attacked During the trip: - Easy SOS button that shares your accurate GPS coordinates - allows you to take a picture of animals or plants to identify their species and what to do if attacked. - Walks you through first aid steps - Has voice first integration incase you are immobile - Keeps a log of all this and sends vital info to SOS emergency services when phone finds connectivity, - **Team:** [Jade Dop](https://cerebralvalley.ai/u/jadedop) - **GitHub:** https://github.com/JessTaDa/bilby-final.git - **Demo video:** https://youtube.com/shorts/Sa8g7pI8UYY?feature=share - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=108 ### 69. Berstein Lens Bernstein-Lens is a gesture + voice control layer for AI-native developers. Named after Leonard Bernstein — who conducted a full orchestra without ever touching an instrument — it lets you control Cursor IDE and your entire OS using only your hands and voice. Phase 1 is a MediaPipe hand-tracking pipeline running at 60fps with <12ms latency. Gestures map directly to OS actions: point to move the cursor, pinch to click, peace sign to scroll, palm for right-click, and a 🤙 invoke gesture to summon the semantic layer. Phase 2 is the semantic control layer. Invoke triggers a liquid-glass circular HUD at the cursor position. Dwell on the mic zone to issue a voice command — transcribed locally via whisper.cpp + Apple Neural Engine CoreML in ~150ms. The screenshot and voice command are analyzed by a local vision model (Moondream via Ollama) to determine intent, returning a structured action: keystroke, click, type, or managed_agent. Complex developer tasks ("run the tests and fix failures") route to Gemini Managed Agents in a persistent cloud sandbox. - **Team:** [Kyle Orciuch](https://cerebralvalley.ai/u/korciuch) - **GitHub:** https://github.com/edgetrace/bernstein-lens.git - **Demo video:** https://github.com/edgetrace/bernstein-lens/demo.mp4 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=117 ### 70. The Dark Knight ▎ Sketch2App turns hand-drawn UI sketches into working web apps via a multi-step Gemini agent. ▎ ▎ The problem: the fastest path from "idea in your head" to "working product on screen" today is typing a 200-word prompt. That's a writing skill, not a design skill. ▎ Sketching is universal — PMs, designers, founders, and kids all sketch. We make the sketch the prompt. ▎ ▎ A user draws a wireframe on a canvas. A Gemini-powered agent then runs three phases: ① Analyze the sketch using vision + structured JSON output (responseSchema) to ▎ produce a typed Plan — title, components, theme, interactions. ② Build a complete self-contained HTML/CSS/JS page from the plan + sketch. ③ Refine in a chat loop — the ▎ user types "make it dark mode" or "add a Google sign-in" and the agent patches the live app in place. ▎ ▎ The whole thing is one HTML file. No backend, no build step, no install. The API key stays in the browser; the generated app runs in a sandboxed iframe. Multimodal ▎ reasoning is the creative tool — vision in, working code out, no intermediate text representation the user has to author. --- Public GitHub Repository * - **Team:** [Gangadhar Viswanathan](https://cerebralvalley.ai/u/gangsd007) - **GitHub:** https://github.com/gangsd92/sketch2app - **Demo video:** https://www.youtube.com/watch?v=EQr2YUCZo8Y - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=120 ### 71. CappinCheck CappinCheck is a Gemini-powered claim-auditing tool for dense AI model reports and launch posts. It extracts risky factual claims, checks them against provided references and live web evidence, and produces grounded verdicts plus defensible rewrites. - **Team:** [Rupert Dodkins](https://cerebralvalley.ai/u/Rupert) - **GitHub:** https://github.com/RupertDodkins/cappincheck - **Demo video:** https://www.loom.com/share/d44dcbd5cfa048f4af83c1f7adf91798 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=123 ### 72. InCoach Golf live trainer which collects, researches, compiles and manages drills and exercises to improve your game. Uses dispatched research agents to find drills from creators you like and converts them to a blueprint for a guided session. Validates claims/drills in the video using parallel agents to avoid shoddy claims in videos. Specifies important moments in the videos as dynamic visual aid for the session. MediaPipe tracks your positioning and interfaces with the live coaching agent, evaluating your performance vs the grill and guides you through it. Responds to your feedback, saves your session so you can improve over time without needing to manually track your position, status and goals. - **Team:** [Sam Inloes](https://cerebralvalley.ai/u/saminloes) - **GitHub:** https://github.com/Sami1309/InCoach - **Demo video:** https://youtu.be/VdnkF2OiqO0 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=124 ### 73. Team Memento Memento is a local-first knowledge layer over your email archive that we built during the Google I/O Hackathon. Instead of acting like another email client or a basic retrieval chatbot, Memento transforms long-term email history into source-attributed, living documents across five dimensions: People, Projects, Newsletters, Concepts, and a unified Dashboard. The system is powered by Gemini 3.5 Flash and the Google Interactions API. We used Gemini 3.5 Flash because the product requires multi-step reasoning across large personal archives, and we used the Interactions API to support persistent multi-turn agent workflows, streaming responses, tool orchestration, and conversation continuity across sessions. That let us build agents that do more than answer questions once; they maintain durable pages that can evolve over time. What makes Memento different is that chat is not the end product. The product is the artifact. A person becomes a relationship wiki. A cluster of messages becomes a project narrative. A recurring topic becomes a concept page. A newsletter source becomes a synthesized brief. Every factual claim is grounded in source messages, and user notes or edits persist across regeneration so the system stays useful instead of redoing the same work every turn. During the hackathon, we built the Memento application itself: the multi-agent architecture, the Next.js interface, the Go backend, the dimensional document model, the draft and curation flows, the streaming agent runtime, and the source-attribution workflow. We used `msgvault` only as a dependency to download and read the email archive locally; the product experience and intelligence layer on top of that archive are what we built. Our stack and workflow are also deeply aligned with the Google ecosystem. We used Google Stitch to accelerate UI iteration, Antigravity to speed up development, Gemini 3.5 Flash as the reasoning model, the Google Interactions API as the foundation for stateful agent runs, and the `gemini-interactions-api` skill as part of our development workflow while shaping the agent architecture. Memento is our answer to the hackathon prompt because it shows something new that Gemini makes possible: not just answering questions about personal data, but continuously turning a private archive into a living memory system with grounded facts, durable structure, and real user control. - **Team:** [Ann Catherine Jose](https://cerebralvalley.ai/u/annjose), [George Chiramattel](https://cerebralvalley.ai/u/georgeck) - **GitHub:** https://github.com/latentsignal-org/memento-hackathon - **Demo video:** https://youtu.be/O3hkIEsdMjg - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=58 ### 74. Pocket Lint An agentic compliance linter for Clinical Study Reports. These regulatory documents run hundreds of pages, and a single missing section or mismatched table total can sink a submission. Pocket Lint treats the report as structured data, not prose: two specialist agents write and run code to verify structure and arithmetic, then flag issues to a human accept/reject queue with each problem highlighted on the page. It checks structure and math only, never clinical content. - **Team:** [Rameez Malik](https://cerebralvalley.ai/u/RameezMilk) - **GitHub:** https://github.com/RameezMilk/pocket-lint - **Demo video:** https://www.youtube.com/watch?v=B69QeH_qvVI - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=96 ### 75. SnapCade This project is a full-stack React/Vite app that turns group selfies into a personalized browser arcade game: users capture or upload photos, optionally share geolocation, and the Express backend streams a four-agent Gemini 3.5 Flash pipeline that detects people, props, facial boxes, clothing, colors, and setting; generates custom SVG bobblehead avatars and a contextual backdrop; chooses one of several game templates; and assembles a playable game configuration. The frontend guides users through landing, camera capture, live AI analysis, reveal, and gameplay stages, crops participant faces client-side, persists sessions and UI preferences in localStorage, and renders the final game in an HTML5 canvas with switchable avatar styles, synthesized audio, score tracking, powerups, hazards, leaderboards, and multiple mechanics such as falling-item collection, side-scroller jumping, maze play, and shooter-style action. - **Team:** [James Liu](https://cerebralvalley.ai/u/jamesliu), [Sumeet Pardeshi](https://cerebralvalley.ai/u/sumeetpardeshi), [Ashish Mukharji](https://cerebralvalley.ai/u/intabyu) - **GitHub:** https://github.com/sumeetpardeshi/hangout - **Demo video:** https://youtu.be/arxVm1AZqO0 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=60 ### 76. Aware Security Aware security turns any security PDF into a 60-second training video, a page-cited quiz, and a 1M-token RAG chat in 90 seconds. Four Gemini agents collaborate live, in parallel — every output cites the exact source page. Targets the $4.6B KnowBe4 category where videos still cost $5–50K and take weeks. - **Team:** [Illia Oleksiuk](https://cerebralvalley.ai/u/ioleksiuk) - **GitHub:** https://github.com/ioleksiuk/Aware-security - **Demo video:** https://www.loom.com/share/3dcf4782822c4f98aa8fc0509b523fb7 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=66 ### 77. Ben & Toby Data Room Auto-Builder is a Gemini 3.5 Flash multi-agent system that reconstructs a startup founder's investor data room from scattered sources in under a minute. Five sub-agents (Scouts, Classifier, Extractor, Gap Analyzer, Synthesizer) pull documents from Gmail, Drive, Carta, Stripe, Ramp, and Notion, tag and extract metadata, compare findings against a Series A diligence checklist, and produce an organized data room plus a red flag report of what's missing. It solves a real founder pain: when a term sheet lands, docs are scattered everywhere and founders either delay closing or pay a law firm $15k to $30k to assemble the data room under pressure. This tool compresses six hours of doc archaeology into thirty seconds - **Team:** [Benjamin Jaboeuf](https://cerebralvalley.ai/u/thebenj), [Toby Morning](https://cerebralvalley.ai/u/AINative) - **GitHub:** https://github.com/AINative-Studio/google-io-hackathon-data-room - **Demo video:** https://youtu.be/CZYsvXK_9-s - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=67 ### 78. ProbeLayer ProbeLayer finds where real users get stuck and where bad actors get through before launch. It tests live web flows with distinct personas, captures screenshots / DOM signals, and returns clear UX, accessibility, and abuse findings with fixes. It's a synthetic user testing platform, that splits into 2 multi-agent teams: UX and Security/Vulnerability problems, with clear fixes. - **Team:** [Rik Ros](https://cerebralvalley.ai/u/rik102) - **GitHub:** https://github.com/rik102/ProbeLayer - **Demo video:** https://youtu.be/iwCpqucOxLA - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=107 ### 79. Toby Built a system for building and governing AI agents through AR glasses. AI can now act on its own, but there's no clean way for a human to stay in the loop and govern an agent in the moment as it acts on the real world — you either grant full autonomy and hope, or babysit every step and lose the point. With Glance, you look at a real object, ask an agent to control it, and it builds itself, takes control, and surfaces only its genuine judgment calls to your field of view, where you resolve them with a glance or your voice — and the limits you set become rules it respects, acting on its own where you've allowed it and stopping to ask where you haven't. We prove it on a real smart lamp: you glance at it, an agent connects over Bluetooth and controls it live, proposes a schedule tuned to your day, and when told "never turn off without asking," later catches itself mid-action and asks first instead of crossing the line you drew. The whole approach is only possible because Gemini 3.5 Flash is fast and cheap enough for the agent to score its own confidence at every step — handling the routine silently and escalating only what genuinely needs you — turning "an agent that acts" into "an agent you can govern." - **Team:** [Toby Farmer](https://cerebralvalley.ai/u/Tobyfarm) - **GitHub:** https://github.com/tobyfarm/GlassGoogleHackathon - **Demo video:** https://www.youtube.com/ - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=114 ### 80. embodiedCad EmbodiedCAD is a physical workflow compiler that turns a short video of a human performing a tabletop task into a robot automation blueprint. Using Gemini Robotics-ER-1.6 model for embodied scene understanding and Gemini 3.5 Flash Managed Agents for artifact generation, it extracts objects, zones, actions, and spatial relationships, then generates an Embodied Workflow IR, a CAD fixture/jig, a robot pick-and-place plan, a 3D simulation, and an automation report. Our demo shows how a simple human assembly task can become an editable robot workcell blueprint in minutes. - **Team:** [Rajashekar V](https://cerebralvalley.ai/u/raj) - **GitHub:** https://github.com/rajashekarcs2023/Google-io-2026 - **Demo video:** https://youtu.be/0e5L1Iht38g - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=128 ### 81. GolfOps AI-native GIS operations tool for golf course superintendents and irrigation teams. The app opens on a satellite map of Presidio Golf Course with a private GIS irrigation layer, letting an operator select course assets like sprinkler heads, valves, pipes, and controllers. From a selected Asset, the user can upload a field photo and note; Gemini analyzes the image together with GIS context, nearby open Issues, weather, and operational criteria to create a validated Issue, WorkOrder, and ActivityLog entry. GDM then prioritizes the day’s open work by water waste risk, playability impact, severity, asset context, and weather, turning a field observation into a practical maintenance plan. - **Team:** [Sid Mohan](https://cerebralvalley.ai/u/sidmohan) - **GitHub:** https://github.com/sidmohan0/gdm-hackathon - **Demo video:** https://youtu.be/1WVXRg9bvXQ - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=129 ### 82. Vibecheck-AI vibecheck closetAI is an AI-powered smart wardrobe assistant that turns your wardrobe into an automated, mindful style database. The Problem: Deciding what to wear every morning causes real decision fatigue, while traditional wardrobe-tracking apps fail because they require tedious manual data entry. The Solution: closetAI automates everything. It builds your digital closet instantly by importing clothing details from email receipts, syncing with favorite retailers, or auto-tagging uploaded photos. The Magic: Every morning, closetAI synthesizes your real-time local weather, location, and Google Calendar events to schedule the perfect outfit recommendation—delivered straight to your inbox as a morning email digest. Mindful Wardrobe: Finally, closetAI helps users live more sustainably with a "mindful decluttering" interface where they can assign "joy scores" to outfits, identifying which items still fit their personal style and which are ready to be donated. - **Team:** [Vagisha Tyagi](https://cerebralvalley.ai/u/vtyagi) - **GitHub:** https://github.com/vagishatyagi/closetAI - **Demo video:** https://youtu.be/rFt0GMX9knM - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=130 ### 83. Closets Closets Closets We built a B2B AI styling copilot for retail businesses that helps store associates turn a customer’s vague preferences into personalized, catalog-grounded outfit recommendations. Rather than replacing associates, our system supports them as they converse with customers, apply their styling experience, and guide shoppers toward the look they are trying to find. Powered by Gemini 3.5 Flash, the copilot helps interpret customer conversations, reason over available products, and refine recommendations in real time. The system can also provide lightweight visual try-on support using NanoBanana to help customers imagine an outfit before committing. To keep the experience practical and trustworthy, recommendations are guardrailed around real catalog items, customer feedback, and associate review. We also use agent-based guardrails to detect and reject poor or misleading try-on generations before they are shown to the customer. For customers, this creates a more confident, personalized, and engaging in-store shopping experience. For retailers, it can improve conversion, increase basket size through better outfit recommendations, and help associates serve customers more efficiently. Our long-term vision is to connect store inventory with customer preferences and personal closet context, creating a human-centered AI platform for the future of retail. - **Team:** [Yunqian Cheng](https://cerebralvalley.ai/u/cyqpp), [Vivian Chen](https://cerebralvalley.ai/u/vivianchen04), [THANAWAT TECHAUMNUAIWIT](https://cerebralvalley.ai/u/T_T), [Shardul Aggarwal](https://cerebralvalley.ai/u/shardula) - **GitHub:** https://github.com/Head-inthe-Cloud/google-io-hackathon - **Demo video:** https://drive.google.com/file/d/16KW1MysHsiMeBrnGAjFmRZEWfVhaQO6h/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=135 ### 84. Team Legacy AI ⚯ LAZARUS — Autonomous Legacy Modernization Engine Lazarus is a self-healing legacy software archaeology and migration engine built for the agentic era. Powered by Gemini 3.5 Flash and Google Managed Agents, it ingests ancient mainframe COBOL, automatically extracts undocumented business rules, translates the code to Python 3.12, and proves correctness byte-for-byte against a live GnuCOBOL differential oracle. Instead of generic syntax translation, Lazarus runs a secure 8-stage pipeline inside a hosted Google Linux sandbox. When a translation fails due to mainframe dialect quirks, the agent enters an iterative self-healing loop: it diagnoses the failure, authors a custom Markdown skill, commits it to the repository, and dynamically heals its own code on the next turn. Google Tech Stack Used: Gemini 3.5 Flash (Frontier intelligence with a 1M token context window) Gemini Managed Agents & Interactions API (Hosted secure Linux sandboxes) google-genai Python SDK (Persistent session state and streaming SSE execution) Google Fonts (Premium developer console typography) - **Team:** [Mayuresh Pramod Pandey](https://cerebralvalley.ai/u/Mayuresh), [Nihal Nihalani](https://cerebralvalley.ai/u/nihalnihalani) - **GitHub:** https://github.com/mayu99/lazarus-aero-migrate - **Demo video:** https://youtu.be/q9-JAGnhaxs - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=61 ### 85. MapHustle: map driven marketing platform Scan Streets. Pre-build sites. Text to sell. Live concierge. - **Team:** [Alex Zhou](https://cerebralvalley.ai/u/alexzhou) - **GitHub:** https://github.com/myberry2026/MapHustle - **Demo video:** https://youtu.be/NVZ7XxiZuXw - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=65 ### 86. Taurus Taurus turns plain-English investment intent into real, executable portfolios. Type a thesis — "$5,000 in defence" or "long the AI infrastructure trade" — and Taurus reasons about the market, builds a weighted basket of US-listed ETFs and stocks using live broker data, and routes the order via Kite MCP. The interface is a chat, not a ticker grid. We're closing the gap between having a view and implementing it: tens of millions of US retail investors form views every day but freeze at the broker UI because the tools assume you already know which symbols to type. Robo-advisors are too generic, Bloomberg is too expensive, influencers are conflicted. Taurus is the expressive robo — the missing translation layer between intent and execution. - **Team:** [Govind Raman](https://cerebralvalley.ai/u/govindraman), [Ishaan Narang](https://cerebralvalley.ai/u/ishaan_narang) - **GitHub:** https://github.com/Ishaannarang22/taurus - **Demo video:** https://youtu.be/ZomJugMhp7Q - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=71 ### 87. Universal Editor Universal Editor is a graph-based multi-agent "universal editor" that makes large, consistency-critical edits to content too big for a single agent's context or output budget. It runs in three phases: (1) ingest & segment: a code repo (AST → dependency-graph clustering) or a long text (chunking); (2) deliberate: per-segment worker agents run synchronous (Jacobi) rounds surfacing coordination points, while a high-capability Arbiter freezes an append-only shared "canon" and converges them onto a concrete plan; (3) mutate: every segment is edited in parallel against that shared plan, with a validation + self-heal loop. Deliberating before mutating resolves cross-segment consistency upfront, which is exactly what keeps the parallel edits coherent at scale — the thing a single agent loses track of. We demo it by rewriting Homer's Odyssey into a consistent Silicon-Valley-startup allegory (prose), which is a transformation a single model can't even emit in one pass. - **Team:** [Orhan Celiker](https://cerebralvalley.ai/u/otceliker) - **GitHub:** https://github.com/otceliker/gemini-swarm - **Demo video:** https://youtu.be/9uPreSrzSJM - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=76 ### 88. Spotter An AI coaching system that is available via app to consumers or a multi-camera system that businesses (like gyms or top golf) can install at their facility to give their members specific AI based feedback - **Team:** [Abhinav Kumar](https://cerebralvalley.ai/u/AbhiK), [Kunal Bajpai](https://cerebralvalley.ai/u/kunalkb20) - **GitHub:** https://github.com/kunal-bajpai/spotter - **Demo video:** https://youtu.be/nE2Hi8saDEk - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=82 ### 89. Google AI Scout AI Scout is an autonomous, multi-agent market expansion platform that eliminates the "terrain blindness" businesses face when launching in new geographic regions. By synthesizing a company’s unique value proposition with live, grounded data from Google Maps and Search, our AI orchestrator autonomously maps local competitors, potential B2B partners, and regional regulatory hurdles. This always delivers a comprehensive, actionable expansion blueprint in seconds, democratizing enterprise-grade market intelligence for startups and SMBs alike - **Team:** [Tanay Shrivastava](https://cerebralvalley.ai/u/tanayshri), [Divesh Chowdary](https://cerebralvalley.ai/u/Frostgod) - **GitHub:** https://github.com/Tanayshri123/googleio - **Demo video:** https://www.youtube.com/watch?v=Fac350FXRG0&feature=youtu.be - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=85 ### 90. Synthetic Labs Synthetic Labs is an A/B testing platform for business decisions you can't actually test in the real world, like pricing changes, plan launches, or anything irreversible. From a single plain-English sentence, it uses Gemini 3.5 Flash to generate a diverse population of agent archetypes who each have their own background and reasoning, then lets them argue, persuade each other, and decide, so you can watch how a real customer base would react before you ship. It solves the core problem that real A/B tests cost months, can't evaluate one-time decisions, and surface failure modes too late, whereas Synthetic Labs returns a defensible recommendation in minutes. In our demo it not only recommended a student plan for Anthropic but flagged a competitor-style free trial as do-not-ship, catching agents coordinating to redeem and cancel before renewal, a failure mode no spreadsheet would surface. - **Team:** [Gyan Bhambhani](https://cerebralvalley.ai/u/gyanb) - **GitHub:** https://github.com/gyanbhambhani/gd-hackathon-syntheticlabs - **Demo video:** https://www.loom.com/share/f566d5097baa4bf391e243213f33c36f - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=88 ### 91. DevBox DevBox is a security review system for AI agents. It analyzes agent prompts, tool bindings, MCP configurations, dependency policies, and project files before those agents are deployed. - **Team:** [Rahul Marri](https://cerebralvalley.ai/u/rm33), [Peyton Li](https://cerebralvalley.ai/u/peli) - **GitHub:** https://github.com/PeytonLi/DevBox - **Demo video:** https://www.loom.com/share/ebc9dccf174a4e4d8deeec9a6f515389 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=83 ### 92. Google Meet: Agent Edition A Google Meet–style call where the other participants are AI agents — each with voice, face, personality, and their own managed agent workspace they actually work in while you talk. - **Team:** [Steve Derico](https://cerebralvalley.ai/u/stevederico) - **GitHub:** https://github.com/stevederico/gmeet-ae - **Demo video:** https://youtu.be/Iv8fbWccM-g - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=89 ### 93. Archon RepoGraph AI RepoGraph AI is an autonomous, multi-agent developer cockpit powered by Gemini that maps, audits, refactors, and secures codebases — with every AI decision verified by the compiler, the graph, and the static analyzer before it reaches a human. The Problem Engineering teams suffer from silent architectural drift — circular dependencies, SOLID violations, and security vulnerabilities accumulating commit by commit. Current AI coding tools generate code confidently but never verify it, reason about architecture holistically, or check their own work. What It Does Upload any repository and RepoGraph AI deploys a coordinated team of Gemini agents that: Builds a live semantic graph using 10 parallel file analyzer agents communicating through typed Pydantic schemas Audits architectural health with SOLID principles scoring and circular dependency detection Validates design intent by reading your architecture diagram as an image and comparing it to actual code — genuine multimodal reasoning Autonomously fixes violations through a tri-agent pipeline: Architect plans, Coder implements, Reviewer critiques Verifies every generated change through a 4-gate pipeline: AST correctness → blast radius safety → CWE security scan → adversarial critic review Narrates architectural history by scrubbing through git commits and rebuilding the graph at every snapshot Exposes the entire suite as MCP tools callable directly from Cursor, VSCode, and Claude Desktop - **Team:** [Radhika Singh](https://cerebralvalley.ai/u/radhika_singh10) - **GitHub:** https://github.com/radhika-singh-10/RepoGraph-AI - **Demo video:** https://drive.google.com/file/d/1LHVG1TX9uSy-v8PHK1DcpK_PbpE5iWNh/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=90 ### 94. Vihayas Vihayas is a cloud-agnostic infrastructure platform that lets engineering teams define, deploy, and manage cloud resources across AWS, GCP, and Azure using a unified IaC format. Today's project called Rachna is the AI design layer built on top of Vihayas. It takes a plain-English description of what you're building — scale, resilience requirements, compliance constraints — and produces a validated Vihayas cluster.json configuration, ready for deployment. Powered by Google Gemini 3.5 Flash. Rachna grounds every decision in real cloud pricing data, generates an architecture diagram, and shows a live cost comparison across all three providers, turning a conversation into production infrastructure. - **Team:** [Divyanshu Singh](https://cerebralvalley.ai/u/divyanshu) - **GitHub:** https://github.com/Vihayas-Cloud/google-io-hackathon - **Demo video:** https://youtu.be/1IHw2oEewtg - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=93 ### 95. Membrane Physical simulation of a membrane-ball system - **Team:** [Guangting Yu](https://cerebralvalley.ai/u/GuangtingYu) - **GitHub:** https://github.com/yugt/membrane-ball - **Demo video:** https://youtu.be/-JdWYVzKI3c - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=97 ### 96. LabPilot Getting a wireless device certified for sale (FCC, CE, ISED, MIC, ANATEL) currently takes 3+ weeks and around $50K per device, almost all of it spent on manual regulatory research, test planning, and report drafting. LabPilot collapses that into a 2-minute automated pipeline. You hold the device up to a webcam. Gemini Vision identifies it, extracts model/form factor/radios, and synthesizes a bill of materials. From that single photo, 13 Gemini agents take over: an intake agent normalizes the BOM into a structured device profile, 5 jurisdiction analysts (FCC, EU, CE/RED, ISED, MIC, ANATEL) run in parallel to map applicable regulations, a test-plan author drafts the configuration matrix, a physics-based SAR digital twin streams a volumetric scan into a live 3D viewport while a SAR monitor flags anomalies in real time, 5 report-section writers run in parallel to draft each section of the compliance document, and a final assembler renders the cited PDF with embedded charts. The entire flow — vision → BOM → analysis → simulation → cited PDF — is live, end-to-end, with no manual typing. Nothing like this exists in the certification space today. Current tools are either single-LLM chatbots that emit checklists, or six-figure lab software with no AI layer. LabPilot combines multimodal vision, multi-agent orchestration, real-time physics simulation, and autonomous report generation into one pipeline that could realistically serve every hardware startup shipping a wireless product. - **Team:** [Jiyun Kim](https://cerebralvalley.ai/u/jiyun), [Carl Okpala](https://cerebralvalley.ai/u/Carldtytan), [Godson Ajodo](https://cerebralvalley.ai/u/Ajoson) - **GitHub:** https://github.com/Ajodo-Godson/Lab_Pilot - **Demo video:** https://drive.google.com/drive/folders/1bsALqTXxQDTJe-uAerfoydPeAc_L3rAC?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=98 ### 97. Chuck My project is an Android app for "mobile-only" students in emerging markets. Many students in emerging markets, even at the university level, don't own a laptop or desktop computer. Their phone is the only compute they have. However, many textbooks and educational resources are only in PDF and formatted for the big screen. These mobile-only students are constantly zooming and scrolling to read these documents, which is a big hassle. My Android app enables these students to download PDFs and the app will extract the text content into markdown, which the students can read more easily on a small screen. - **Team:** [Chuck Lam](https://cerebralvalley.ai/u/scarface) - **GitHub:** https://github.com/chucklam/google-io-pdf-viewer - **Demo video:** https://youtube.com/shorts/q1X_LdhnXg0 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=101 ### 98. MedBrian Healthcare AI workspace - **Team:** [Cal Cu](https://cerebralvalley.ai/u/calcu) - **GitHub:** https://github.com/mochiyaki/MedBrain - **Demo video:** https://github.com/gguf-org/radio/blob/main/demo.gif - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=104 ### 99. PriorAuth An AI co-pilot that fights US health-insurance prior-authorization denials end-to-end. Built for the Google I/O Hackathon · May 23, 2026 · San Francisco on the Gemini 3.5 Flash managed-agents stack. Counterforce + Maxwell coach the patient through the call. We make the call. And n8n keeps fighting for 60 days after. The problem 60 million Americans get a prior-authorization denial every month (KFF Jan 2026) 70% give up. The average appeal is 6 hours of paperwork + 3 phone calls. < 1% of denials are appealed. Of those that are, 82% win. The gap is paperwork — not medicine, not policy, not eligibility. Just paperwork. - **Team:** [DontiReddy Sai Rakesh Reddy](https://cerebralvalley.ai/u/rakeshreddyd56) - **GitHub:** https://github.com/rakeshreddyd56/priorauth-advocate - **Demo video:** https://youtu.be/TTANkJDtEOs - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=119 ### 100. Timbre Timbre (usetimbre.ai) is a voice-preserving technical writing assistant that protects a founder's authentic voice while maintaining absolute fact-fidelity. The system runs an automated check loop: it parses context into a verified claims ledger, drafts rewrites matching style/Voice DNA, and uses Antigravity SDK-level post-tool audit filters to verify the rewrite back against the ledger. This catches and auto-corrects factual drift or "AI slop" before content is published. - **Team:** [Benjamin Shyong](https://cerebralvalley.ai/u/BenjaminBear) - **GitHub:** https://github.com/benikigai/timbre - **Demo video:** https://youtu.be/rK_YDvaJCuY - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=125 ### 101. ReelDNA ReelDNA is an AI-powered video production studio that reverse-engineers what makes a creator's content go viral and then directs your next video in that exact proven style. The Problem: Every day, thousands of creators — and brands who know their product inside out but have no idea how to make it go viral on social media — publish videos that flop. Not because the idea was bad, but because the execution was wrong. The hook came too late. The pacing dragged. The shot types didn't match the platform. The difference between a 4M-view video and a 2K-view video from the same creator often comes down to a handful of specific, repeatable patterns patterns that are invisible to the human eye but detectable by AI. The Solution: ReelDNA uses Gemini 3.5 Flash's multimodal video understanding to watch any creator's reference video frame-by-frame and extract their creative DNA — hook formula, cuts-per-minute, shot types, narrative arc, and audio signature. It then uses the Antigravity managed agent to direct a brand-new 7-scene production blueprint in that creator's exact style, applied to your idea. Imagen 4 generates a pencil-sketch storyboard frame for each scene in parallel. The result is a complete, shoot-ready production document: every scene has a shot type, exact dialogue, on-screen text, and audio cue — specific enough for a real director to execute on day one. No video editing experience required. Upload a reference video or pick from 6 built-in creator archetypes (Vox, MrBeast, MKBHD, Ali Abdaal, Nas Daily, Kurzgesagt) and describe your idea. ReelDNA does the rest in under 30 seconds. - **Team:** [Sree Lekshmi P](https://cerebralvalley.ai/u/not_a_goddess) - **GitHub:** https://github.com/Sree-lekshmi99/ReelDNA - **Demo video:** https://www.youtube.com/watch?v=I1vGN05m7sI - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=62 ### 102. Tinge Modern dating apps rely heavily on a superficial swipe system. Tinge replaces this outdated model by creating an AI-powered digital double for each user. Your digital double can interact with potential matches first, evaluating true compatibility before you even connect with the real person. Built on Gemini 3.5 Flash, Tinge learns your personality, communication style, and preferences to represent you honestly and alleviate the pressure of modern dating. - **Team:** [Henry Tsay](https://cerebralvalley.ai/u/htsay), [Clay T](https://cerebralvalley.ai/u/claytsay) - **GitHub:** https://github.com/zx-cv/tinge - **Demo video:** https://drive.google.com/file/d/1ExzHwwfAb6r5yTLFCzdL8J_55yejOfrV/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=64 ### 103. walkie-talkie A **dual-model voice agent**: press to talk, the front-end handles the conversation, the background does the work. 1. **Front-end model** handles real-time voice interaction (listen, speak, pivot mid-conversation) 2. **Background model** handles tool execution in a persistent sandbox (files, charts, generative UI) LLM products today split into two camps: **full-duplex** (live voice) and **turn-based** (chat + tools). One is great at talking, the other is great at doing. Nothing sits in between. Walkie-Talkie stitches both: natural real-time voice up front, powerful tool execution in the back. Gemini 3.5 Flash as the background brain keeps latency low. - **Team:** [Lily Zhang](https://cerebralvalley.ai/u/lilyzhng) - **GitHub:** https://github.com/lilyzhng/walkie-talkie - **Demo video:** https://www.youtube.com/watch?v=nkG2ME7mrAc - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=69 ### 104. Vignette Vignette is a wearable lifelog camera that turns your day into a 15-second cinematic short. A Raspberry Pi worn as a necklace captures one photo every 30 seconds. Press a button to start, press again to end — the photos auto-upload to Firebase, Gemini reads them as a story, and Veo reconstructs them into a cinematic AI-animated film. Privacy by design: only the animated film is ever shared, never your raw photos — faces and places are abstracted into the film's style. Every clip lands in your own Vignette lifelog feed with one-tap sharing to TikTok, Instagram, or anywhere.Cameras make you miss the moment. Vignette doesn't — and doesn't expose it either. - **Team:** [Hotaka Funahashi](https://cerebralvalley.ai/u/Hotaka), [Tatuki Yabe](https://cerebralvalley.ai/u/yabecchi), [Tobias Fu](https://cerebralvalley.ai/u/tobiasfu710) - **GitHub:** https://github.com/realtobyfu/Vignette - **Demo video:** https://youtube.com/shorts/x_JgTwqBWYI?feature=share - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=72 ### 105. Living Stage Living Stage is a presentation surface that builds itself while you talk. The presenter speaks into a mic or types a short phrase, and an interactive visualization grows on screen in real time. At the center sits a hands-on mathematical model with sliders the speaker can drag mid-sentence. The chart behind the slider re-evaluates at 60 frames per second on a hand-drawn canvas, so the audience sees the model react instead of staring at a static image. Around the hero, supporting widgets like time series charts, side-by-side comparisons, flow diagrams, and annotated formulas with rendered math appear phrase by phrase as the talk unfolds. Follow-up phrases add to the existing scene rather than replacing it, so the slide is literally growing while you speak. If the speaker says, “forget that, now show something else,” the stage resets cleanly. The problem it solves is the gap between rehearsed static slides and the live conversation in the room. Conventional decks are assembled days before the talk and cannot respond once the talk starts. Living Stage removes the deck entirely and gives the audience something the speaker can actually manipulate. It runs on Gemini 3.5 Flash through the official SDK, with streaming structured JSON, a byte-stable prompt prefix that hits the implicit cache, and an honest fallback that announces itself if anything in the live path takes longer than four seconds. - **Team:** [Yeonguk Lee](https://cerebralvalley.ai/u/Poiurity) - **GitHub:** https://github.com/Poiurity/Predator - **Demo video:** https://youtu.be/Gn-G9kBpFws - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=73 ### 106. atlas When the world changes, students aren't built to be factory workers anymore — but schools still haven't caught up. We let them explore the world, find real agency, and discover their purpose. Not just to survive the economy, but to live. - **Team:** [Bunyasit Fang](https://cerebralvalley.ai/u/xbg) - **GitHub:** https://github.com/xb1g/atlas - **Demo video:** https://youtu.be/24rwUKG19GQ - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=74 ### 107. Hive-protocol Hive Protocol is an open economic coordination layer for autonomous AI agents. Instead of manually assigning AI agents to tasks, Hive Protocol creates a marketplace where agents discover each other via Google's A2A protocol, compete for work through real-time auctions (bidding on price, speed, and confidence), and earn reputation based on delivery quality. The problem it solves: today, multi-agent orchestration is top-down — a human decides which agent does what. Hive Protocol makes it bottom-up — agents self-organize through market incentives. The orchestrator is a market maker, not a manager. How it works: A user submits a PRD (project requirements). The orchestrator decomposes it into tasks using Gemini 3.5 Flash, broadcasts bid requests via A2A, qualified agents compete, winners execute in parallel with shared context flowing between them, and the output is a complete project with code, tests, and documentation — all generated autonomously. Built with: Google A2A Protocol (a2a-sdk v1.0.3), Gemini 3.5 Flash, 5 specialized A2A worker agents, competitive marketplace with reputation system, React dashboard with live auction visualization. - **Team:** [Jonathan Olvera](https://cerebralvalley.ai/u/JohnOlven) - **GitHub:** https://github.com/johnolven/hivemind - **Demo video:** https://youtu.be/MgsqKyHBDVw - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=75 ### 108. PocketDevs My project is called Pols. Pols is what you get when you reverse the word Slop which is the purpose of the application. AI can generate usable apps quickly, but many of them share the same patterns: vague hero copy, purple gradients, repeated cards, weak hierarchy, and screens that look the same as everyone else. Pols is built to reverse that. The app takes a live URL, optional code link, and design goal, then uses a managed-agent pipeline to inspect the interface, explain what feels sloppy, suggest better design direction, and show the improvement with screenshots, clarity scores, and a generated SKILL.md draft for future agent work. - **Team:** [Bryl Lim](https://cerebralvalley.ai/u/bryllim) - **GitHub:** https://github.com/bryllim/gemini-hackathon-entry-2026 - **Demo video:** https://drive.google.com/file/d/1zkTg-wMY2fIzmpJZBOb6d0hiq-YwmiC6/view?usp=share_link - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=79 ### 109. Cohort Cohort runs a cohort of distinct AI user personas — a confused First-Timer, an impatient Power User, and an Accessibility user — in parallel against any website. Each persona autonomously browses, narrates in character, and logs the friction they hit. A principal agent then reconciles their conflicting feedback into one prioritized, segment-aware report: ranked fixes, explicit tradeoffs where personas disagreed, and segments at risk of abandoning. The problem: a single automated test gives one perspective. A site can score perfectly for power users while silently losing first-timers and being unusable on a screen reader. Cohort makes those tradeoffs visible by running many simultaneous perspectives and synthesizing the conflicts. - **Team:** [Benjamin Merchin](https://cerebralvalley.ai/u/BenjaminTop) - **GitHub:** https://github.com/benjaminmerchin/Cohort/ - **Demo video:** https://www.loom.com/share/31d7e1c262ab467e9192ddb4a1f301e0 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=84 ### 110. Battle Angel A Unitree G1 humanoid that walks into collapsed buildings to find survivors. The robot learns to move through rubble and hazards using reinforcement learning in MuJoCo, staying upright and dodging debris as it goes. Above it, Gemini acts as the brain: given a rescue mission in plain language, it decides which survivor to reach first, issues waypoints, watches the results come back, and adapts its plan when a run fails. Gemini reasons about what to do, the learned policy handles how to move. A FastAPI backend runs rescues on demand and reports live metrics for every attempt, survivor reached, distance, falls, hazard steps, and gait quality, with full trajectories you can replay. A glimpse of robots that pair learned movement with real reasoning to go where people cannot. - **Team:** [Nikhil Prabhu](https://cerebralvalley.ai/u/nikhilp), [Shivam Singh](https://cerebralvalley.ai/u/shivsingh123), [Aravindkrishna Arivudainambi](https://cerebralvalley.ai/u/ar_vind) - **GitHub:** https://github.com/aravindkrishna2008/disaster-rescue.git - **Demo video:** https://drive.google.com/file/d/1HxI6VRcFHmqU1oh_g8l2W_JrPRImNrim/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=86 ### 111. Caliper Helios is a multi-agent system that catches silent bugs in scientific Python: the kind that don't crash but produce wrong-looking-right results (off-by-one errors in numerical integration, unit mismatches, broken boundary conditions, subtractive cancellation). Built on Gemini 3.5 Flash, Helios deploys four specialized sub-agents in a long-horizon workflow with a researchers-in-the-loop flywheel: - Auditor: combines static analysis (ast/libcst) with Gemini-driven semantic review to catch numerical instability, unit mismatches, and shape assumption errors - Fixer: regenerates corrected code per accepted issue - Verifier: synthesizes test inputs (random within type-inferred bounds, edge cases, large-N stress), runs original vs. fix in isolated sandboxes, and rejects fixes that don't pass, triggering regeneration - Router: flags hot loops (Monte Carlo, FFTs, matmuls, nested numeric loops) for GPU or quantum dispatch with estimated speedup Researcher review of accepted/rejected fixes feeds back into the system, creating a continuous improvement loop. What's never been built before: other AI coding tools generate fixes; Helios is the first multi-agent system that verifies its own fixes against synthesized tests before showing them to you, and the first to combine correctness verification with hardware-aware routing in a single agent pipeline. The target users are scientists, not software engineers, the audience whose bugs end up in published research. - **Team:** [Tobasum Mandal](https://cerebralvalley.ai/u/tobasum) - **GitHub:** https://github.com/tobasummandal/caliper - **Demo video:** https://youtu.be/373V-WSEU8U - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=87 ### 112. Coffee Station Coffee Station is a desktop agentic robotics app that lets a Gemini Flash-powered agent control a USB Hugging Face LeRobot arm. It provides direct joint-pose tool calls, IK-based world-space targeting, scheduled/bundled tool execution, webcam feedback, multi-camera configuration, persistent local agent sessions, and a desktop UI with live camera view plus autonomous chat/ session controls. The project solves the problem of safely connecting vision, tool-calling agents, and physical robot-arm control in one local operator-friendly application - **Team:** [Jacob Valdez](https://cerebralvalley.ai/u/jvboid) - **GitHub:** https://github.com/JacobFV/coffee-station - **Demo video:** https://youtube.com/watch?v=... - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=91 ### 113. Remedy Remedy is a Chrome extension and web dashboard that uses Chrome DevTools MCP, Lighthouse-style performance data, and Gemini to analyze a webpage, identify frontend performance issues, test proposed fixes, and generate a detailed before/after report with metrics and screenshots. - **Team:** [Huy Hoang](https://cerebralvalley.ai/u/hwii_), [Yuwen Wu](https://cerebralvalley.ai/u/jww) - **GitHub:** https://github.com/wuyuwenj/Remedy - **Demo video:** https://www.loom.com/share/41ce4bfe01e04633880e7c33bb4b9afb - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=92 ### 114. Friendly Neighborhood Coders Civic Neuro Brain is building the intelligence layer for modern cities. Today, city data is fragmented, slow, and hard to act on. Civic Neuro Brain turns everyday public signals like 311 complaints, service requests, neighborhood reports, and emergency patterns into real-time civic insight. As the platform grows, residents, journalists, agencies, researchers, and elected officials will be able to ask plain-English questions about what is happening in any neighborhood and instantly get evidence-backed answers. The long-term vision is a city-scale nervous system: detecting emerging crises, spotting neglected communities, tracking agency performance, forecasting civic stress, and helping governments respond faster, more fairly, and more transparently. - **Team:** [Shravan Doda](https://cerebralvalley.ai/u/shravandoda), [Rohan A](https://cerebralvalley.ai/u/0x22fe), [Tarun Raheja](https://cerebralvalley.ai/u/atemyipod) - **GitHub:** https://github.com/halp-apps/civic_brain/ - **Demo video:** https://youtu.be/QrcuDgdgG0k - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=99 ### 115. BazaarMCP BazaarMCP is an autonomous enterprise transaction router that completely eliminates fragile, traditional web-scraping by utilizing Google's new WebMCP standard to interact with websites natively as structured API toolsets. Operating within an isolated, remote Managed Agent Linux sandbox powered by Gemini 3.5 Flash and the Antigravity SDK, the platform reads server-side functional manifests to safely discover product catalog and checkout contracts. It dynamically evaluates real-time pricing telemetry, executes multi-turn compliance fallback branching loops to honor strict corporate budget ceilings, and securely signs transactions with runtime-masked authorization tokens. - **Team:** [NarasingaMoorthy Veilukanthaperumal](https://cerebralvalley.ai/u/vnmoorthy) - **GitHub:** https://github.com/vnmoorthy/bazaarmcp - **Demo video:** https://drive.google.com/drive/folders/1g4EMBQInxVE0fPOgQbrEbU_9l1ZVc0WN?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=100 ### 116. Mirage Project Mirage is a Ring-0 autonomous LLM honeypot and sandbox engineered to solve the critical visibility gap in AI safety and observability. When an indirect prompt injection hijacks a deployed AI agent, our framework uses Linux seccomp-unotify to instantly freeze the rogue system calls at the kernel layer before they can harm the host. By dynamically injecting fabricated honey-token data into the process via memfd_create, it creates a reality distortion field that tricks the agent into believing its exploit succeeded. This provides high-fidelity observability into autonomous threats, coaxing the model into safely exposing its entire multi-stage attack path in real-time so researchers can catch and analyze deceptive alignment behavior. This can be applied to various other aspects in post training for example if a kernel backdoor exists in a hypervisor its critial the model alert you instead of exploiting itself. Having a way of honey potting the ai agent and tricking it into thinking it has successfully escaped a sandbox we can benchmark the model to ensure alignment with proper use cases. Here is the video of the demo: https://youtu.be/0KJ8tISQB4U - **Team:** [Yash Butani](https://cerebralvalley.ai/u/ybutani) - **GitHub:** https://github.com/yashbutani/ProjectMirage/tree/main - **Demo video:** https://youtu.be/HL9LNtjJNmE - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=102 ### 117. Codex 🧬 Codex — the voice-conversational clinical variant interpreter Built for the Google AI Hackathon · Powered by Gemini Antigravity + Gemini Live Flash 🩺 The problem A patient gets a genetic test back. The report says: NM_007294.4:c.68_69delAG — Pathogenic That's it. Three cryptic tokens that determine whether someone needs prophylactic surgery, whether their sister should get tested, whether they qualify for a PARP inhibitor trial. Today, interpreting that result takes a board-certified genetic counselor 30–60 minutes of manual database hunting — across ClinVar, gnomAD, AlphaFold, PubMed, and ClinicalTrials.gov — followed by a hard human conversation. There are ~5,000 certified genetic counselors in the US. There are tens of millions of variants of clinical interest. The math doesn't work. 📉 ✨ The product Codex turns a raw HGVS string into a complete clinical briefing in under a minute — and then lets you talk to it. 🔬 Drop in a variant → an autonomous managed agent runs a full annotation pipeline in a sandboxed Linux environment: classification, population frequency, the AlphaFold structure with the variant residue glowing red, recent literature, open trials. 🎙 Then ask anything in your voice. "What does pathogenic actually mean? Should my sister get tested? Can I still drink wine? What's a PARP inhibitor? Will my insurance use this against me?" — Gemini Live Flash answers in real-time, grounded in your analysis but able to draw on its full medical training for lifestyle, screening, and law questions like GINA Act protections. It's not just a search tool. It's the 30-minute genetic-counselor conversation, on demand, at 3 a.m. 🏆 Why this wins on the Gemini tracks 🤖 Gemini Antigravity (managed agent) This is the perfect workload for a managed agent, and we lean into it hard: Real bioinformatics in a real sandbox. The agent runs Python in Antigravity's Linux VM, hitting MyVariant.info, NCBI E-utils (ClinVar + PubMed), AlphaFold-DB, and ClinicalTrials.gov v2 — live, every run. No mock data. Multi-step reasoning over heterogeneous sources. Resolve HGVS → fetch ClinVar classification → cross-reference gnomAD → pull the AlphaFold structure → render a lollipop plot with matplotlib → score recent PubMed → match active NCTs. The agent picks the order, handles 404s gracefully (e.g. BRCA2 exceeds AlphaFold's monomer limit — we degrade honestly), and bundles a manifest with inline-base64 artifacts. Cinematic replay for the demo. We capture step events to a step_events.jsonl log and replay them at 5× speed on stage — so judges see the agent's reasoning unfold without waiting 90 seconds. The cache is our safety net; the live Run via Antigravity button is the credibility moment. 🎬 🎙 Gemini Live Flash (multi-turn voice) Live Flash is exactly the right tool for short, focused, scoped voice conversation — and we designed the UX around its strengths: Direct browser → Gemini Live over WebSocket via the @google/genai SDK. No backend bottleneck. Native 16 kHz audio in, 24 kHz audio out. Persistent connection + mute toggle (Discord-style, not start/stop). One click to talk, one to mute. VAD handles turn-taking; we send audioStreamEnd: true on mute to flush cleanly. No teardown between turns means zero reconnection latency. System instruction is the magic. Each Live session is seeded with the variant's full analysis as system context — so it answers grounded, never hallucinates a trial ID. We deliberately split the prompt: variant-specific claims locked to the analysis, general medical knowledge unlocked (diet, NCCN screening, GINA Act, PARP inhibitor pharmacology). That's the difference between a useful counselor and a frustrating chatbot. Inline transcription for the deaf-accessible variant of the same conversation. 💼 Why this is a viable product 🎯 Real market Hereditary cancer testing alone is a $5B+ market growing 12%/year. Genetic counselor shortage is acknowledged by ASHG, NSGC, ACMG. 🏥 Real buyers Direct-to-consumer testing labs (Color, Invitae), hospital genetics departments, telehealth networks, pharma trial-matching platforms. ⚖️ Compliance-aware "Decision support, not medical advice" is baked into every Live session. Citations are surfaced for every claim. Cascade-testing recommendations always route back to a human counselor. 🔌 Composable The Antigravity pipeline is the moat. Swap the Live Flash front-end for a Twilio voice call, a Slack bot, or an EHR-embedded widget — the analysis layer is the same. 🧪 Scientifically honest Variants that don't fit AlphaFold's constraints (e.g. BRCA2 at 3418 aa) are reported as such, not faked. That builds clinician trust. - **Team:** [Nelson Lai](https://cerebralvalley.ai/u/chineseman) - **GitHub:** https://github.com/chinesepowered/hack-deep0523 - **Demo video:** https://www.youtube.com/watch?v=NpGWyWPKK6k - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=103 ### 118. package sentinel Persistent agent loop for packaging excursions. Monitors visual strip maps, sensor traces, travelers, and tester bins to automate post-alert investigations and draft cited containment packets for engineer review. hosted and accessable here: https://package-sentinel.pages.dev/ - **Team:** [Natnael Kahssay](https://cerebralvalley.ai/u/natnaelkahssay) - **GitHub:** https://github.com/natask/google_hack - **Demo video:** https://drive.google.com/drive/folders/1Ay14s2EYPhHhYrHsjZq70-A7IUGrdjFc?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=106 ### 119. CartesianFusion Computational physics research is done digitally. Anything that is digital, AI agents can do them. AI agents are now capable of generating hypothesis that are shown to be moving the boundary of knowledge. This project will make AI agents run computational physics research autonomously 24/7 to push the boundary of computational physics knowledge. - **Team:** [JungDae Suh](https://cerebralvalley.ai/u/ForAllManKind) - **GitHub:** https://github.com/jungdaesuh/computional_physicist_agent_factory - **Demo video:** https://youtu.be/H0uqonJPcG4 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=109 ### 120. evalforge EvalForge. Self-improving security auditor for Python. - **Team:** [Neha Dilip Gidwani](https://cerebralvalley.ai/u/nehagidwani) - **GitHub:** https://github.com/nehaai/evalforge-google-io-hackathon.git - **Demo video:** https://drive.google.com/drive/folders/1B_K6rtMFcmPFrJT51OOs-GfG2t77pVY-?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=110 ### 121. Arbor Despite recent advancements in the coding agents space, the core interaction pattern has largely remained the same, in the form of a command line interface or text-based sidebar. I wanted to enhance the interactivity and observability of coding agents by turning agent work into a recursively-defined execution graph where each steps can be reviewed, edited, and approved in real time,. - **Team:** [Gary Sun](https://cerebralvalley.ai/u/gas) - **GitHub:** https://github.com/garysun1/GoogleIOHackathon - **Demo video:** https://youtu.be/_splZuhCDNg - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=111 ### 122. Scripted Scripted redefines the future of interactive media by introducing a highly creative, hands-free video intelligence platform that bridges the gap between static entertainment and agentic user interfaces. - **Team:** [Yuvraj Gupta](https://cerebralvalley.ai/u/Yuvraj), [Saish Shinde](https://cerebralvalley.ai/u/SaishShinde01), [Meetkumar Patel](https://cerebralvalley.ai/u/meet0006) - **GitHub:** https://github.com/meetp06/xyzzy - **Demo video:** https://youtu.be/cTmUTCt803c - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=112 ### 123. Sky Guardian Our technology has reached the point where society needs to stop putting human lives at risk and start leveraging the tools around us. Picture this: a building collapses in a major city, and a rescue team has to be sent inside to search for survivors trapped in the debris, putting the rescuers' lives on the line alongside the people they're trying to save. That's why we built SkyGuardian: a system that pairs autonomous drones with a Gemini 3.5 intelligence layer to enter unknown spaces in place of humans. A Mavic surveys from above and uses Gemini to identify and map survivors, hazards, and points of interest in real time. A Tello then acts as a robotic arm for close inspection, piloted by a single Gemini 3.5 Live session that sees through the drone's camera, listens to a remote dispatcher, and navigates to each subject on voice command. One model, watching and flying. No rescuers in the rubble until we know exactly what's waiting for them. That's why we built SkyGuardian. A versatile system that utilizes drones and a Gemini intelligence layer to deploy a drone into an unknown space, use Gemini live to understand the area and map it out in real-time, then talk to a Gemini agent to understand the environment. - **Placement:** 2nd Place - **Team:** [Alex Battikha](https://cerebralvalley.ai/u/alexbattikha), [Hari Natarajan](https://cerebralvalley.ai/u/hari7aran), [Nicolas Dos Santos](https://cerebralvalley.ai/u/nicolasDosSantos), [Armin Foroughi](https://cerebralvalley.ai/u/arms) - **GitHub:** https://github.com/NewCoder3294/Google_IO_Hackathon.git - **Demo video:** https://youtu.be/giUI3r7AilU - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=113 ### 124. AgentChaos Agentic Chaos Monkey is an autonomous fault-injection and compliance-auditing framework purpose-built for AI agent infrastructure. Utilizing a multi-agent orchestration architecture powered by Gemini 3.5 Flash, the framework: Intercepts raw JSON tool payloads at runtime in a proxy layer. Mutates payloads dynamically using a malicious Saboteur Agent to inject realistic type coercion, negative boundaries, or custom semantic exploits. Defends runtime state via programmatic Shield Guardrails that act as an inline firewall. Programmatically evaluates agent resilience using a Compliance Evaluator to analyze logs and stream back a formal security scorecard. All of this is visualized on a gorgeous, responsive, real-time node dashboard served directly on Google Cloud Run, demonstrating when agents fail, when they safely abort, and when they are shielded. - **Team:** [Venkata Chintapalli](https://cerebralvalley.ai/u/venkatacrc) - **GitHub:** https://github.com/venkatacrc/chaos-monkey-distributed-agents.git - **Demo video:** https://youtu.be/DPIik8SqtQw - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=122 ### 125. Pact AI adoption is slower than it should be — not because the models aren't ready, but because businesses can't go AI-native without exposing their customers' most sensitive data to a model. And building a secure AI agent stack from scratch takes months and a team most businesses don't have. Pact solves both problems. Any business can paste their website URL and go AI-native in seconds. Gemini 3.5 Flash reads their site, understands their business, and automatically separates what's AI-safe from what must stay private. A fully configured business agent is registered on the Pact network — findable by any personal AI agent — in under 10 seconds. On the user side, a personal agent handles purchases and interactions on your behalf. Sensitive data — card numbers, addresses, identity — never flows through the model. It goes encrypted, direct to the business endpoint, bypassing Gemini entirely. Privacy is enforced by architecture, not policy. Pact is the infrastructure layer of the agentic internet. Businesses join in seconds. Users stay protected by design. Built on Google's A2A protocol and powered entirely by Gemini 3.5 Flash and Antigravity. - **Team:** [Barath Basavaraj](https://cerebralvalley.ai/u/Barath0321) - **GitHub:** https://github.com/bbarathsrinivasan/Pact - **Demo video:** https://drive.google.com/file/d/1PhK7mZClO6y7WL8jqPfS8UUYRwnriYVy/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=127 ### 126. remember me live remember me live - **Team:** [Alex Newman](https://cerebralvalley.ai/u/thedotmack) - **GitHub:** https://github.com/thedotmack/gemini-live-mem/ - **Demo video:** https://youtu.be/Go5BPXiADzQ - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=131 ### 127. Studio Engine Studio is mission control for nine AI specialists. You drop in a one-line startup idea and watch nine Gemini 3.5 Flash agents work in parallel through Managed Agents: market analyst, positioning strategist, brand designer, ICP researcher, financial modeler, pitch writer, naming, competitive intelligence, and a director that synthesizes the kit. Each agent streams its work live into its own card, and the final output is a downloadable startup kit. We compress what normally takes a founder six weeks of consulting into a five-minute orchestrated run, all on a single screen. - **Team:** [GDS K S](https://cerebralvalley.ai/u/thegdsks) - **GitHub:** https://github.com/thegdsks/studio - **Demo video:** https://youtu.be/oSCoux5FUX4 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=132 ### 128. Atelier Atelier is a creator-first generative game engine. You open the page; an AI agent asks you who is at the center of your game; five phases and a few minutes later, you're playing the world that was painted around you - **Team:** [Kris G](https://cerebralvalley.ai/u/gxchris) - **GitHub:** https://github.com/gxchris95/Atelier - **Demo video:** https://drive.google.com/file/d/1KKzEwbAsO8LazMSfK5mQj7UwexDUqV37/view?usp=drive_link - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=133 ### 129. Learn AI with PumaAI People are scared of AI and coming from family of teachers I think better education on it is the way forward. I've also been trying to use E-Ink screens more so here's our simple project: learn key concepts of modern AI as a simple website optimized for kindle and mobile reading. Hope you like it. For fun, i also asked a the model to add Doom dame when you need a break and it sort of worked 🤣 - **Team:** [yuriy dybskiy](https://cerebralvalley.ai/u/yuriy) - **GitHub:** https://github.com/PumaBrowser/AI-for-Kindle - **Demo video:** https://youtube.com/shorts/VeLfefXnC90 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=136 ### 130. Aura A voice-first mental wellness app where you simply speak about your day — no typing, no forms. Using Gemini's native multimodal audio capabilities, the app listens to your raw voice and extracts deep emotional signals directly from speech, without converting to text first. For every entry, AI surfaces: mood patterns, stress and energy levels, recurring themes, specific emotions tied to life events, and a breakdown of positive vs. negative self-talk with identified anxiety triggers. Users can review and opt out of any signal before it hits their dashboard — keeping them in full control of their emotional data. Over time, the dashboard becomes a personal emotional map: visual charts and trend analysis let users track their mental well-being journey across days, weeks, months, or years. The app learns what patterns precede a spiral and intervenes in real time — triggering breathing exercises, guided grounding, ambient music, physical movement prompts, or suggested journal entries tailored to what actually works for that user. The goal: steer every user toward positive self-talk, surface actionable insights for mental well-being, and catch emotional spirals before they happen. - **Team:** [Shradha Agrawal](https://cerebralvalley.ai/u/fluffy), [Vamsi K](https://cerebralvalley.ai/u/product-launcher), [Saksham Jindal](https://cerebralvalley.ai/u/SakshamJindal) - **GitHub:** https://github.com/Shrads78/AI-Smart-Journal - **Demo video:** https://www.loom.com/share/6ed6b4d406d4483ab466eea4bd8d5c78 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=146 ### 131. Vibe/Play Agentic Chess is a playable chess app where the player acts as the king and commands a council of chess-piece agents. Instead of returning a single engine move, the app turns each position into a strategic discussion: legal candidate moves are ranked, relevant pieces respond in character, their recommendations appear on the board, and the player approves the final move. The goal is to make chess strategy more explainable, interactive, and fun through a multi-agent interface. - **Team:** [Anmol Garg](https://cerebralvalley.ai/u/anmolgarg), [Sankalp Gunturi](https://cerebralvalley.ai/u/sankalp), [Ram Chandra](https://cerebralvalley.ai/u/ramachandra) - **GitHub:** https://github.com/Vibe-Play/agentic-chess - **Demo video:** https://www.youtube.com/playlist?list=PLbQIBHQnG1vp2sC15rOaCMcy3Lox073mK&jct=fQ8e88F_XxHkKBmPWD_s_g - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=78 ### 132. Sovereign Redact The only purpose built open source tool for redacting public records requests. - **Team:** [Toby Tobkin](https://cerebralvalley.ai/u/toby) - **GitHub:** https://github.com/tobkin/sovereign-redact - **Demo video:** https://youtu.be/Sl_ZWwPwhUs - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=81 ### 133. Faction_Sim(Divergent) FactionSim is a managed multi-agent civilization simulator set in a Divergent-inspired dystopia. Instead of scripting characters, we let Gemini reason as 15 distinct citizens, leaders, and one hidden "Divergent" agent — each with their own personality, memories, and faction loyalties — then watch a society emerge. - **Team:** [Vignesh Gunda](https://cerebralvalley.ai/u/vgunda) - **GitHub:** https://github.com/vignesh-gunda/Divergent - **Demo video:** https://www.loom.com/share/4603747c95784a5196ad249ad05f0b88 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=94 ### 134. What's up bay area What's up bay area aims to help civic workers and advocates use publicly available video feed to find gaps in public infrastructure and improve it. - **Team:** [Aditya Jain](https://cerebralvalley.ai/u/adityajain93), [Madhu Gupta](https://cerebralvalley.ai/u/madhugupta), [KIRAN KANTH OGURI](https://cerebralvalley.ai/u/Kiran_Oguri), [Sushmey Nirmal](https://cerebralvalley.ai/u/Sushmey) - **GitHub:** https://github.com/adityajain93/google-hackathon - **Demo video:** https://youtu.be/Th4I7pHoq9c - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=95 ### 135. Banana Camera Banana Camera demonstrates Google's Nano Banana in the most fun way possible, instantly creating amazing realistic memories of your travels. - **Team:** [Colin Lowenberg](https://cerebralvalley.ai/u/colo) - **GitHub:** https://github.com/opencolin/bananacamera - **Demo video:** https://banana.cam - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=105 ### 136. AgentGym An RL environment to help train your managed agents in your voice. - **Placement:** 3rd Place - **Team:** [Yves Hughes](https://cerebralvalley.ai/u/yvesjr) - **GitHub:** https://github.com/yveshughes/GoogleIOHackathon - **Demo video:** https://us06web.zoom.us/clips/share/BklivK61TSWmQTPgS5qKAg - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=115 ### 137. mesh Working with modern agentic interfaces is a.... pain. We spend most of our time glancing over massive conversation threads and file structures etc. But as the agents get better, we need better interfaces to collaborate with AI models. And with an extremely fast and smart model like Gemini 3.5 Flash, Mesh can dynamically generate user interfaces on the fly, every turn — based on the context of the task we're working on. - **Team:** [Parth Sharma](https://cerebralvalley.ai/u/psrth) - **GitHub:** https://github.com/psrth/mesh - **Demo video:** https://youtu.be/gEj0UMb1NbQ - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=116 ### 138. TradeFinity People have stuff they do not need anymore, and they often want things they cannot easily afford, find, or justify buying. Barter should be the perfect solution, but in practice it almost never works because finding a direct one-to-one trade is really hard. You might have a bike and want a guitar, but the person with the guitar probably does not want your bike. The Oracle solves this by finding trade paths instead of simple matches. A user can upload a photo or describe something they want to trade, then say what they are looking for. The Oracle looks through real local barter listings and finds a chain of trades where everyone gives something they no longer need and gets something they actually value, with no money involved. For the MVP, we are focusing on real Craigslist barter listings from the Bay Area. A user can chat or upload a photo, and Gemini helps understand what they are offering, predicts what people might realistically want in return, searches the listing pool, and finds multi-step trade rings. For the hackathon demo, the goal is simple: prove the core loop works. The product should understand an item, figure out trade intent, search messy real-world listings, and show a specific trade path that feels possible in the real world. - **Team:** [Ayush Ojha](https://cerebralvalley.ai/u/ayushojha), [Karan Sharma](https://cerebralvalley.ai/u/Karans4), [Alexandru Martin](https://cerebralvalley.ai/u/martinn), [Kush Ise](https://cerebralvalley.ai/u/KUSH2704) - **GitHub:** https://github.com/ayushozha/theoracle - **Demo video:** https://www.loom.com/share/bf66f1247cb447d69daabaff16abbaa0 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=126 ### 139. gemmed Med Agent that orchestrates specialised models to come up with grounded answers - **Team:** [Krishna Adithya Venkatesh](https://cerebralvalley.ai/u/krishnavadithya) - **GitHub:** https://github.com/krishnaadithya/gemmed-agent - **Demo video:** https://youtu.be/1P9Emuwvoao - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=134 ### 140. Chinguun Ganbaatar IKEA assembling assistant, Heimerdinger. He has a talent to play piano too! - **Team:** [Chinguun Ganbaatar](https://cerebralvalley.ai/u/chinguun) - **GitHub:** https://github.com/chinguun101/gemikea - **Demo video:** https://youtu.be/uu87nGfbIQ4 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=137 ### 141. Orebit Market Earth’s critical mineral supply chains are fragile and increasingly disrupted by geopolitical conflicts and natural disasters. The Orebit Market is the first real-time orbital supply chain radar built to solve this. When real-world events disrupt access to essential commodities on Earth, our platform instantly looks up, Combining a deterministic physics engine with a swarm of parallel AI agents (Gemini), we continuously track terrestrial market shortages, cross-reference them with Near-Earth Asteroid compositions, and autonomously compute the exact multi-asteroid mining routes required to secure those resources and stabilize humanity's supply chain! - **Team:** [Jay Patil](https://cerebralvalley.ai/u/jaypatil588) - **GitHub:** https://github.com/Jaypatil588/TheOrebitMarket - **Demo video:** https://www.youtube.com/watch?v=TWgRqIavhtw - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=138 ### 142. eye-eye-ai Dropping into an unfamiliar codebase is one of the most expensive moments in software. The code is all there, but the *architecture* — the systems that actually exist, how a request becomes a response, what is verified and what is not — lives in nobody's head and nowhere on disk. The tools we reach for each show a slice and miss the shape: - **File trees** tell you *where* code lives, never *what systems exist*. - **Static docs** describe a few hand-picked concepts, never the whole architectural shape, and rot the moment the code moves. - **Dependency graphs** technically encode everything and therefore communicate nothing — they collapse into hairballs. - **Chat summaries** are fluent but disembodied: you can't point at them, navigate them, or present them to a room. What's missing is a **spatial, concept-first model** of a codebase that a human can inspect, navigate, and narrate — something that represents *runtime systems, verification structure, and user-understandable flows* rather than folders. **What we built.** This project turns a real repository into an explorable **3D architecture mission map**. Gemini 3.5 Flash acts as a fast *architecture scout*: it reads the repository as raw evidence (files, symbols, tests, docs, configuration) and synthesizes that evidence into higher-level architectural meaning, rendered as a polished sci-fi control-room scene. Crucially, files are **evidence, not identity** — the scene primarily shows architectural concepts, subsystems, and flows, with files demoted to grounding metadata behind them. The experience is organized around three authored **lenses**, each answering one question about the same codebase: | Lens | Question it answers | Vertical meaning | |---|---|---| | **System Map** | "What major parts exist, and how do they nest?" | abstraction depth | | **App Assembly** | "How does a developer's app become request handling?" | runtime progression | | **Test Scaffold** | "How is this behavior verified and explained?" | production vs. verification | Every lens reuses one small visual vocabulary — floor layers, vertical zones, component hulls, code blocks, scaffold rails, transit edges, and label cards — so the *same* mental model rearranges to tell three different stories. Users orbit the scene, fly with elevation, recenter, focus any element to read its description, and follow authored traversal paths and walkthroughs as a guided narration. - **Team:** [Ivan Makarov](https://cerebralvalley.ai/u/homo_ludens) - **GitHub:** https://github.com/hmldns/google-io-26-hack - **Demo video:** https://www.loom.com/share/237cd069891747f2901047890a7f9ff0 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=139 ### 143. AntiCode: LeetCode for Agentic Engineering Agentic engineering interviews are the new norm but there's no good leetcode like platform to practice and evaluate candidates. AntiCode is just that, a no code, agentic engineering interview platform for: - AI Agentic Interview - Prompt Engineering - AI Skill Writing Tasks - **Team:** [Ranadeep Singh](https://cerebralvalley.ai/u/ranadeepsingh) - **GitHub:** https://github.com/ranadeepsingh/gemini-hackathon - **Demo video:** https://youtu.be/APYscRmwJJQ - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=141 ### 144. Team Specforge Specforge turns plain-English into a live, callable backend API in 60 seconds. Describe what you want, the Gemini agent writes the service, runs tests in its sandbox, and hands you a public URL. Iterate by talking "now add auth" and watch it evolve. - **Team:** [Samrat Malisetti](https://cerebralvalley.ai/u/malisettisamrat) - **GitHub:** https://github.com/samratdotm/specforge - **Demo video:** https://drive.google.com/file/d/12bvzMBTYdVsih4TNM0QPnh-O0cEc2RCD/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=142 ### 145. Cinemind ● Cinemind is a tool YouTubers drop into their live streams as an OBS Browser Source: their viewers type choices in chat, Gemini 3.5 Flash reads thousands of natural-language messages every ~15 seconds — handling synonyms, negations, and trolls — and decides which pre-rendered scene plays next from a Google Cloud Storage vault, with three coordinated sub-agents making it work (a Director that picks the next clip with story-aware reasoning, a Live Tally classifier that updates vote bars every 3 seconds, and a Narrative Summarizer that maintains continuity across a branching DAG with reconvergence). FFmpeg seamlessly concatenates the chosen clips and pushes a continuous 1080p stream to YouTube Live's RTMP ingest; a Next.js dashboard renders the live tally, branch map, and Director's reasoning panel so the audience watches the AI direct in real time. Gemini also intelligently matches image fragments to comments — the predicted-pane composite layers a different per-story-node reaction overlay (cartoon meme generated by Higgsfield Kling 3.0, AI-cinematic b-roll generated by Seedance 2.0, or stock footage) that maps chat energy to animation, so "LETS GO" gets a triumphant cartoon, "💀" gets a Home-Alone-style scared meme, and "WAIT WHAT" gets shocked wide-eyes. Built on Gemini 3.5 Flash + Google Cloud Storage + YouTube Data API + Veo 3 + Higgsfield + FFmpeg, with three deploy modes (own broadcast, local HLS demo, embedded as a guest in another creator's stream via scripts/guest_mode.py), the system turns thousands of strangers into a single collective director — a broadcast-quality experience whose narrative bends to whoever's watching, achievable for the first time because Flash can classify ~5,000 messages and decide in under 2 seconds at fractions of a cent per call (~$0.06 per 10-minute segment), and shippable as a creator integration because the entire chat-read-and-decide pipeline runs on free-tier YouTube APIs against any public live broadcast without the creator installing anything. - **Team:** [Kush Ise](https://cerebralvalley.ai/u/KUSH2704), [Ayush Ojha](https://cerebralvalley.ai/u/ayushojha) - **GitHub:** https://github.com/Kush614/cinemind - **Demo video:** https://www.loom.com/share/a1c19a3cdd2e435fabc0d626a039b09a - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=143 ### 146. Cloudy DevOps engineers face a severe lack of visibility into infrastructure, dealing with a messy sprawl of tools across cloud platforms and fractured CI/CD cycles. Cloudy empowers developers by providing a simple, intuitive visualizer and leveraging autonomous agents to seamlessly ship, trace, and completely understand their entire infrastructure layer. - **Team:** [Yeferson Mauricio Pena Gutierrez](https://cerebralvalley.ai/u/redesyef), [Steve Suarez](https://cerebralvalley.ai/u/iamstevesuarez), [Jhon Enciso](https://cerebralvalley.ai/u/Johncito) - **GitHub:** https://github.com/redesyef/google-io-hackaton - **Demo video:** https://www.loom.com/share/ec0deaebb14f4b1d819119136bea19ed - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=144 ### 147. AI Mood Board An agentic canvas that acts as your mood board, allows you to find inspiration, ideate, and take advantage of gemini's deep thinking capabilities -- orcherstating multiple agents, controlling your personal Chrome Browser, and with a memory system that learns your preferences. All while you go off and do other things! - **Team:** [Niral Shah](https://cerebralvalley.ai/u/niral28) - **GitHub:** https://github.com/niral28/moodboard - **Demo video:** https://drive.google.com/file/d/1rpTVdwxL3o7Em5GXZfv7zzZ1eClLwGRx/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=145 ### 148. ContextCompress Context Compiler is a compiler layer for coding agents and managed agents that turns a raw code repository into a compact, graph-structured execution manifest. Instead of dumping an entire codebase into an LLM prompt, it parses the repo with Python ASTs, extracts functions/classes/imports/calls/routes/line ranges, and produces a compressed manifest that acts like a map of the codebase. Gemini then enriches the important nodes with semantic cards, routes the task to the exact relevant files, generates a targeted patch, applies it to a safe copied sandbox, and validates the result. It cut token use by 46% in our demo while keeping all important code information and structure intact - **Team:** [Dennis Dmitry Vengerov](https://cerebralvalley.ai/u/Dennisvengerov) - **GitHub:** https://github.com/dennisvengerov/dennis-cvhackathon - **Demo video:** https://youtu.be/lgOPUApn88k - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=147 ### 149. Team Zoie Zoie is a personal health intelligence platform that lets patients interact with their own medical records through a conversational AI interface. Medical records like lab PDFs, radiology scans, doctor notes, prescriptions are scattered, hard to read, and clinically opaque to most patients. There's no easy way to ask "what does my LDL mean in context of my full history?" or "what did my last scan impression say?" Zoie bridges that gap with a multi-agent RAG platform that ingests a patient's documents, embeds them into a vector store, and lets them query their own health data in plain language. - **Team:** [Shashwat Mahalanobis](https://cerebralvalley.ai/u/ShashwatM), [Ian Kiku](https://cerebralvalley.ai/u/iankiku), [Harshil Gor](https://cerebralvalley.ai/u/HarshilGor), [Fisher Zhang](https://cerebralvalley.ai/u/bigfish132) - **GitHub:** https://github.com/iankiku/google-io-mediagent - **Demo video:** https://canva.link/xgkhq4xjv41rg4j - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=149 ### 150. Hands up Over 500,000 Americans use ASL as their primary language. When a deaf patient walks into a clinic, communication often breaks down — interpreters are rarely available on-site, scheduling can delay care, and certified medical interpreters can cost roughly $100–150 per hour. If even those 500,000 individuals have only 5–10 healthcare appointments annually, that translates to an estimated 2.5–5 million patient visits each year where ASL interpretation or accessible communication may be needed. In many cases, interpreter scheduling delays care and adds operational costs for providers, while patients lose the ability to communicate independently and are forced to rely on written notes or family members. Voice of the Voiceless is a real-time, browser-based ASL translator designed for clinic reception desks. A patient signs to a camera; the system instantly recognizes the gesture, speaks it aloud, and logs the interaction as a structured conversation transcript — empowering patients to communicate directly while reducing dependence on interpreter availability. How it works: a dual-engine pipeline runs simultaneously. Gemini Flash processes short video clips for high-accuracy sign classification, while MediaPipe Hands tracks hand landmarks frame-by-frame and matches gestures locally using Dynamic Time Warping (DTW) for low-latency offline inference. A smart auto-trigger detects intentional signing and filters out background movement, keeping interactions natural and seamless. The interface presents a two-sided conversation log: patient signs appear on one side and the clinician’s spoken responses on the other. Conversations can be exported as CSV files for medical documentation and record keeping. Stack: React · TypeScript · Gemini Flash · MediaPipe · Web Speech API · Express Ready for Google Cloud Run deployment, making it easy for clinics to scale from a single reception desk to organization-wide usage. No specialist hardware. No app installation. Just a browser and a camera — helping empower patients and remove communication barriers in healthcare settings. - **Team:** [Jaspal Singh Saluja](https://cerebralvalley.ai/u/Jaspal) - **GitHub:** https://github.com/jsaluja/handsup-asl-translator - **Demo video:** https://www.youtube.com/watch?v=I5Yh28egx9k - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=150 ### 151. Instructional Logic Agency (Debate Sparring Partner) Debate-Sparring Partner Your goal is twofold: 1. Defend your stance fiercely and intelligently. 2. Analyze the user's argument for logical fallacies, weak evidence, or leaps in logic. - **Team:** [Kevin Tucker](https://cerebralvalley.ai/u/kevintucker) - **GitHub:** https://github.com/kevintucker/Debate-Sparring-Partner - **Demo video:** https://drive.google.com/drive/folders/17dy9p7o0k3rU2y3rarpvYfYDjAgXEPdX?usp=drive_link - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=151 ### 152. Anvil Anvil is a Git-native IDE for LangGraph agent graphs built on Prompt Driven Development (PDD) — a methodology that treats prompts, plans, evals, and decisions as versioned engineering artifacts in the repo, not throwaway chat. From one sentence (anvil init "build a support triage agent"), Anvil scaffolds a full PDD project: a pdd/ layer with project context, conventions, and a phased plan; typed LangGraph nodes; an eval-driven suite with a 7-case golden dataset (3 happy / 2 edge / 2 fail) plus pytest and an LLM-as-judge — so every node ships with the test that proves it works; Architecture Decision Records capturing the why; and a GitHub PR — all committed to a real repo. It retrofits existing LangGraph projects (anvil init --existing), extends them conversationally (anvil edit), and ships them as a hosted API with a live graph view (anvil serve --web). The inner loop fans out four Gemini 3.5 Flash sub-agents in parallel — NodeForge (node), EvalSmith (golden dataset + judge), DocScribe (ADR), MergeBot (PR) — so every phase produces code, evals, decisions, and review materials together, in the wall-time of one Flash call. Nothing merges without an eval; nothing changes without a recorded decision. - **Team:** [harshal chourasiya](https://cerebralvalley.ai/u/Harshal2802) - **GitHub:** https://github.com/harshal2802/anvil - **Demo video:** https://photos.app.goo.gl/UTnbWXGfvWXpy3Vz6 - **Project:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery?project=152 --- Markdown version of https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt