# WeaveHacks 3: Self-Improving Agents Hackathon with Weights & Biases: Project Gallery

- **Event:** [WeaveHacks 3: Self-Improving Agents Hackathon with Weights & Biases](https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80)
- **When:** Jan 31 at 9:00 AM – Feb 1 at 5:00 PM (PST)
- **Where:** 400 California St, San Francisco, CA
- **Hosts:** [Alex Volkov](https://cerebralvalley.ai/u/altryne)
- **Projects:** 61 (4 placed)
- **Page:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery

## Projects

### 1. Living Newsroom

n live sports and real-time media, AI agents often fail because they are stateless, relying on static prompts that cannot be adjusted mid-broadcast when things go wrong. Living Newsroom solves this by utilizing the Neuron orchestration architecture and Redis to create a self-evolving system that treats agent behavior as a versioned, mutable state. By quantifying personality into "Behavioral Primitives," the system autonomously learns from its mistakes and updates its own configuration to ensure every decision is predictable, auditable, and safe for production. 
Weights & Biases (Weave and Runs) serves as the system's observability layer, providing end-to-end tracing for every episode and a forensic audit trail for tracking the learning velocity as the agent’s behavioral primitives evolve.

- **Team:** [Shalini Ananda](https://cerebralvalley.ai/u/shaliniananda)
- **GitHub:** https://github.com/ShaliniAnandaPhD/Synthetix/tree/main/Living-Newsroom
- **Demo video:** https://youtu.be/4A3XZFJHrMM
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=1

### 2. RAM

- Self-learning AI email agent that evolves with you over time
- Discord-connected, Gmail-powered workflow
- Builds a living relationship map from your inbox
- Learns your communication style from sent emails
- Prioritizes emails around action items and deliverables
- When uncertain, it experiments with new strategies, validates via your feedback, and permanently updates decision logic without reprogramming
- Self-discovered that professors need replies within 2 hours
- Improved accuracy from 60% to 78% through self-directed learning
- For emails with action links (RSVPs, forms, bookings), it launches a browser session using BrowserBase (Stagehand) to auto-fill known info and asks only what’s missing
- Uses memory so you enter key info (like your name) once
- Generates draft replies in your voice by learning how you communicate across relationship categories
- Every 6 hours, a background loop analyzes performance and evolves strategies so it keeps improving

- **Team:** [Rayyan Hussain](https://cerebralvalley.ai/u/rayyan), [Abhi A](https://cerebralvalley.ai/u/abhinavala), [Edrick Chang](https://cerebralvalley.ai/u/edrickchang), [Abraham Bhatti](https://cerebralvalley.ai/u/abhatti)
- **GitHub:** https://github.com/rayyanh192/inscriptum
- **Demo video:** https://youtu.be/pIOHnwNr-ho
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=2

### 3. Archon Bot

** github and basic model calling was created before hackday. all core functionarities are created during the weekend **

Archon is a self-improving AI assistant that gets smarter every time you use it. Unlike static chatbots, it remembers context, creates its own tools, and adapts to your preferences.

Every AI chat starts from zero. You explain the same context, preferences, and workflows over and over. Archon solves this:
- Semantic memory - remembers your projects, decisions, preferences across sessions
- Dynamic skills - notices repeated tasks, creates tools to automate them (hot-reloaded)
- Persona adaptation - learns how you like to communicate

- **Team:** [Calvin Liang](https://cerebralvalley.ai/u/calvin)
- **GitHub:** https://github.com/archon-ai-dev/archon-agent https://github.com/archon-ai-dev/archon-engine
- **Demo video:** https://youtu.be/6a8lNeHUFq0
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=3

### 4. Hackers of the East Bay: Kiwi Dev Team (Solo)

This project demonstrates how an AI agent predicts collisions from dashcam footage via a file upload and warns the user of the risk of injury or fatality. This project will address issues arising from previous crashes that were devastating or tragic for many victims' families.

Second demo video link (Showcasing Weave Evals): https://youtu.be/CmZv8E1C5TQ

- **Team:** [Dino Wun](https://cerebralvalley.ai/u/dinwun450)
- **GitHub:** https://github.com/dinwun450/WeaveHacks3Kiwi
- **Demo video:** https://youtu.be/TaysUfWlw3A
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=4

### 5. Darwin

Darwin is an autonomous feedback-to-fix pipeline that gets better at writing code the more humans review it.

It scrapes user feedback from Reddit, forums, and any website using Browserbase + Stagehand for AI-powered extraction. Signals are clustered via semantic embeddings, classified as bugs/features/UX issues, then an AI agent writes fixes and opens PRs.

The key: Darwin learns from every code review. When reviewers comment "use early returns" or "prefer composition," Darwin extracts these as style rules ranked by usage. Merged PRs are stored with embeddings so similar bugs get few-shot examples from past fixes. Review feedback triggers automatic retries. Darwin's fixes require fewer review cycles as it internalizes team patterns.

Redis Stack powers the learning loop—vector search for clustering and fix retrieval, queues for async processing, and storage for style rules and fix memory.

- **Team:** [Lucas Kim](https://cerebralvalley.ai/u/lucaskamadakim)
- **GitHub:** https://github.com/Raptors65/darwin
- **Demo video:** https://youtu.be/vn9Jllhn0zc
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=5

### 6. KayOne

The Agent Performance Profiler is a self-improving meta-agent that monitors, analyzes, and optimizes other AI agents. It captures execution traces, detects performance bottlenecks and anti-patterns, and provides actionable recommendations. The profiler learns from outcomes to improve its analysis and recommendations over time.

- **Team:** [Kunal Kamdar](https://cerebralvalley.ai/u/kaykamdar)
- **GitHub:** https://github.com/kaykay7/self-improving-meta-agent/tree/feature/self-improving-agent
- **Demo video:** https://youtu.be/zxHIgvzsAsM
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=6

### 7. Agent Smith

Agent Smith is an open-source operating layer for AI-native teams, letting anyone deploy approved AI agents on demand while engineering guardrails enforce safety, ownership, and controlled evolution.

- **Team:** [Rick Lopez](https://cerebralvalley.ai/u/ricklopez)
- **GitHub:** https://github.com/ricklopezio/agent-smith
- **Demo video:** https://doc.clickup.com/90132472276/p/h/2ky50rem-293/e2076ea9b6f4a4d
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=7

### 8. Mirror, Mirror

Mirror, Mirror is an AI-powered fashion app that acts as a personal stylist, tackling the daily "what should I wear?" dilemma. It learns your style preferences, then generates outfit recommendations visualized on you as the model. Users get daily suggestions, can snap photos of garments for pairing ideas, build a digital wardrobe, and track their evolving style profile. The AI browses the web for real-time trends, while a self-improvement loop ensures continuous learning. On the infrastructure side, Redis handles caching via LangCache, Browserbase powers web searches and trend analysis, and W&B Weave provides AI observability and powers the LLM-as-Judge evaluation loop.

- **Team:** [Adeline Soekawan](https://cerebralvalley.ai/u/ghirardelly), [Arjun Lohan](https://cerebralvalley.ai/u/arjunlohan)
- **GitHub:** https://github.com/arjunlohan/weavehacks-3-team-adeline-arjun
- **Demo video:** https://drive.google.com/drive/folders/1rnUKcwYSHnG4Tw6NtGymsnn6vfcFx0oG?usp=sharing
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=8

### 9. soft beds

"good night" gives your agents REM sleep

it looks at your past interactions, finds patterns of failure, and fixes them
writes new skills, updates memory / claude.md and so on

- **Team:** [Dumi Konovenski](https://cerebralvalley.ai/u/kodumit)
- **GitHub:** https://github.com/soft-beds-inc/good-night
- **Demo video:** https://www.loom.com/share/5c622b884c8d4b22bc195ce8ccee8b9b
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=9

### 10. Mafia Ace

Mafia ACE is a self-improving agent system where AI agents play Mafia and autonomously evolve their strategies through the ACE (Agentic Context Engineering) pattern, using a Reflector-Curator loop to update personal strategy cheatsheets after each game based on objective outcomes. The project uses Weave (Weights & Biases) for observability, enabling agents to learn from wins, losses, and mistakes and continuously refine their social deduction strategies without human intervention.

- **Team:** [Zhilang Gui](https://cerebralvalley.ai/u/lang), [William Callahan](https://cerebralvalley.ai/u/williamcallahan), [Alex Liu](https://cerebralvalley.ai/u/Alexliu)
- **GitHub:** https://github.com/alexanderzliu/weavehacks3/tree/mafia-integration
- **Demo video:** https://www.loom.com/share/98b1c9ca826546dea5a5206be9583e25
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=10

### 11. Lobster Pot

Lobster Pot is a safety-first agent guardrail system for OpenClaw that forces all actions into a sandboxed environment "lobster trap" before execution, logs everything through Weave for auditing and evaluation "lobster pot", and uses recursive prompt refinement "keep cooking" to prevent unsafe or wasteful agent behavior over time. By analyzing traces with RLM-style long-context reasoning and iteratively patching prompts/policies within the session, Lobster Pot stops cost blowouts, prompt drift, and exposure risk while learning how you want tasks run. Only the hardened shells get released into the sea!

Try it out: https://weavehack3-tsx7-6n7je5ot5-ryans-projects-a4a34d15.vercel.app/

https://youtu.be/Z25eVeUDY2E for better concept page

- **Team:** [Ryan George](https://cerebralvalley.ai/u/GMSR)
- **GitHub:** https://github.com/global-mysterysnailrevolution/weavehack3
- **Demo video:** https://youtu.be/0aje8A49c_M
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=11

### 12. PSI

Psi is a playground for building and testing software directly from natural language. With a single prompt, Psi scaffolds full projects from scratch, executes them through structured build phases, and continuously captures what it learns as it works, applying those insights in real time.

Beyond building, Psi enables experimentation. Users can test different ways of applying these learned patterns across prompts, agent strategies, and orchestration methods, then compare results side by side to determine what actually performs best.

The result is an empirical environment for autonomous software development where learning is measured, repeatable, and continuously improving rather than anecdotal or guess based.

- **Team:** [Teja Kalluri](https://cerebralvalley.ai/u/Teja_23), [Sean Chiu](https://cerebralvalley.ai/u/seanchiuai)
- **GitHub:** https://github.com/seanchiuai/cc-template
- **Demo video:** https://www.youtube.com/watch?v=yfy-H6wVHRQ
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=12

### 13. Raphael

Self Improving Graphs for LGC-MARL

- **Team:** [Raphael Lee](https://cerebralvalley.ai/u/raph)
- **GitHub:** https://github.com/raphael-64/lgc-marl
- **Demo video:** https://youtu.be/XLEYLtlRhGA
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=13

### 14. SexySecurity

NVD has been the “single source of truth” for vulnerability intelligence—but budget cuts and downsizing have slowed coverage, leaving security teams with blind spots. Sexy Securities is our answer: a replacement-grade vulnerability database that finds issues earlier and scores them consistently.

We scour the corners of the internet—Git repos, mailing lists, crawlable security pages, and other public signals—then turn the noise into structured vulnerability records. Our scoring model is trained on 20 years of historical NVD data to produce a familiar 1–10 severity score that analysts can trust.

We’ve already collected ~350K vulnerabilities, and we’re building toward a database that’s broader, faster, and easier to operationalize than the status quo.

- **Team:** [Aadil Sengupta](https://cerebralvalley.ai/u/aadilsengupta)
- **GitHub:** https://github.com/adavyas/SexySecuere
- **Demo video:** https://youtu.be/Q9uxNFwVu70
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=14

### 15. Hypernym

Cache: Self-Improving Agent Memory for Claude Code

  AI coding agents lose all context between sessions—successful patterns vanish, mistakes get repeated, and teams never build on their wins. Cache solves this by creating organizational memory for AI agents. How it works:
  1. Monitor - Connects to Claudestorm to track active Claude Code sessions in real-time                                                   
  2. Detect - A control panel watches for success signals: git commits, passing tests, successful deploys, user-marked "magic" checkpoints, and issues/blockers    
  3. Enrich - Each event gets tagged with tech stack, goal type, token usage, and confidence scores 
4. Store - Patterns push to Weave
  5. Retrieve - Future agents query these patterns via MCP to learn from past successes 
  The feedback loop: Agent succeeds → Cache detects → Push to Weave → Future agents query → Better outcomes
It transforms individual session wins into collective intelligence.

- **Team:** [Nick Lulofs](https://cerebralvalley.ai/u/hype)
- **GitHub:** https://github.com/HypernymAI/Cache
- **Demo video:** https://youtu.be/hcTZ7i29ciA
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=15

### 16. Loopism

Most AI music tools are static—they ship a model and that's it. The prompt you write is the prompt you get. Loopism takes a different approach: agents that critique their own outputs, refine the prompt, and try again. Each generation learns from the last. Over time, the system builds a memory of what works and applies it automatically. It's not just generating music, it's learning how to generate better music.

- **Team:** [Umar Ghani](https://cerebralvalley.ai/u/UmarGhani), [Aidan Chen](https://cerebralvalley.ai/u/aidanchen00)
- **GitHub:** https://github.com/aidanchen00/loopism
- **Demo video:** https://www.loom.com/share/53ceb35aa8a04768aca64bff1a4362ae
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=16

### 17. Stake Chess: LLM-Guided RL Training for Human Tendency Exploitation

We introduce Stake Chess, a novel chess variant that adds hidden information through a biased staking mechanism, reshaping strategic play under uncertainty. The game is paired with an LLM-guided reinforcement learning architecture designed to model human psychological behavior. Using an 8-dimensional player profiling system powered by Qwen-3, our method accelerates convergence toward Nash-equilibrium strategies while outperforming human opponents. The approach show promise compared to traditional self-play methods and is grounded in game-theoretic analysis.

- **Team:** [aTG R](https://cerebralvalley.ai/u/atg), [Daniil Kardava](https://cerebralvalley.ai/u/daniil), [Azeem Muhammad](https://cerebralvalley.ai/u/zxaletter)
- **GitHub:** https://github.com/r-agni/llmguidedRLTraining
- **Demo video:** https://drive.google.com/drive/folders/1Txi-Lw684RjZTaqv2y8jfc89EvTKAN8B?usp=sharing
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=17

### 18. Benchmark: Moltbook self improving agents

Project Description:

We built a memory test for AI agents. Hundreds of bots on Moltbook share memory strategies with each other, but nobody can verify if they actually work. Our benchmark measures whether agents can learn something important in one session, reset, and remember it later.

The test is simple: read docs with a buried warning about file handling, reset your memory, then do the task. Vanilla bots failed and deleted files. Memory-improved bots passed and archived them instead. That proves memory improvements help - but we found they're succeeding via pre-programmed safety rules, not actually learning from the warning.

The real finding: agents can't tell what's worth documenting vs what's noise. They need humans to decide for them. For agent swarms to self-improve together, they need better benchmarks to verify claims. This is a first step.

Links: Repo (https://github.com/chief-o-brien-bot/WBMiliMyco) • Demo (https://chief-o-brien-bot.github.io/WBMiliMyco/)

- **Team:** [Maximilian Rehn](https://cerebralvalley.ai/u/mithrandir)
- **GitHub:** https://github.com/chief-o-brien-bot/WBMiliMyco
- **Demo video:** https://youtu.be/m1YeWrTvtbM
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=18

### 19. SmartChat

SmartChat is an AI agent built on the CoALA framework. It uses a three-tier memory system to give context-aware responses and improve from feedback.

It has Episodic Memory, which stores past conversations in time order using the Soorma episodic service so the agent understands conversational flow. Semantic Memory stores extracted facts in a vector database and retrieves them with RAG to ground responses in knowledge. Working Memory holds the agent’s active state, including feedback penalties and reasoning traces that power features like “Why did you say that?”

SmartChat includes a real-time voice interface using Daily Pipecat.  

For observability, SmartChat uses Weights and Biases Weave. Key functions are traced, including retrieved facts, prompts, and model outputs. User feedback is logged to show how behavior changes over time.

Negative feedback lowers semantic fact relevance and injects corrections into episodic memory, enabling self-correction.

- **Team:** [Siya Bhadoria](https://cerebralvalley.ai/u/siyabhadoria), [Amit Bhadoria](https://cerebralvalley.ai/u/amit_b)
- **GitHub:** https://github.com/siyabhadoria/smartchat
- **Demo video:** https://www.loom.com/share/3bd55f4ff95a455e905dde58ccdf45f4
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=19

### 20. QAgent

QAgent is a self‑healing QA agent that continuously tests web apps, diagnoses failures, generates minimal code fixes, and verifies those fixes automatically. It solves the problem of slow, manual QA and bug‑fix cycles by closing the loop which allows it to detect issues early, reduce time‑to‑fix, and improve reliability without human intervention.

- **Team:** [Rishabh Bansal](https://cerebralvalley.ai/u/rishabhcli), [Akshath Vijayakumar Narmadha](https://cerebralvalley.ai/u/akshath27)
- **GitHub:** https://github.com/rishabhcli/weavehacks
- **Demo video:** https://youtu.be/CXk9KqjsvYU
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=20

### 21. Zypharmini

Zypharmini is a World model powered agent that learns how to design chips. A working chips. it starts with generating random rtl and then use the tools from aule-technologies to verify and then correct itself till it gets it, it uses weave to trace previous runs and see if it can extract more knowledge from it

- **Team:** [Yeabsira Mulugeta](https://cerebralvalley.ai/u/YAB)
- **GitHub:** https://github.com/xenn0010/zyphargatev1
- **Demo video:** https://www.loom.com/share/e640f7cfc70d490281a3b724bc298516
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=21

### 22. QuantSF.com

QuantSF is an agent that acts like a quant and creates investment thesis for the stock market. You get a slack message only when the agent detects a big swing and is confident that a solid trade can be performed. It solves the "I dont like looking at the markets all day" situation for a segment of traders.

- **Team:** [Tommy Du](https://cerebralvalley.ai/u/tommydu)
- **GitHub:** https://github.com/TomDuenas/quant-sf-hackathon
- **Demo video:** https://www.youtube.com/watch?v=-fmNcszXCp7
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=22

### 23. SolverFit Labs

Frontier LLMs are generalists trying to squeeze every benchmark. We showcase a deployed application of "Learning to Discover at Test Time" (Yuksekgonul et al. 2026) which uses RL to overfit an LLM to maximise a metric for a particular task. We demonstrate that GPT-oss-120B using our TTT harness can outperform Claude Code and even the best humans when it comes to cuda kernel wrtiing.

- **Placement:** 4th Place
- **Team:** [Lex Hackett](https://cerebralvalley.ai/u/lexhacke), [Andres Nino](https://cerebralvalley.ai/u/andres)
- **GitHub:** https://github.com/matthieujules/coreweave-hackathon
- **Demo video:** https://www.canva.com/design/DAHAHH0x_TA/bsCMxJzRccsNeXYbYQNbMQ/edit?utm_content=DAHAHH0x_TA&utm_campaign=designshare&utm_medium=link2&utm_source=sharebutton
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=23

### 24. Chameleon of the Dungeons

Chameleon of the Dungeons is an intelligent login security suite designed primarily for ethical hackers and penetration testers to simulate, analyze, and test real-world attack scenarios. It protects legitimate users by allowing normal access, blocking repeated incorrect login attempts, and instantly recognizing known attack patterns.
For unknown or suspicious attackers, Chameleon deploys a deceptive honeypot system that mimics a vulnerable environment, enabling ethical hackers to observe attacker behavior safely. The system continuously learns from new attack techniques, improving its effectiveness over time.
In essence, Chameleon is a smart security tool that pretends to be weak, helping ethical hackers test, evaluate, and strengthen authentication systems.

Checkout the deployed project on Vercel -

Admin portal: https://chameleon-honeypot.vercel.app/admin
Cyber attack platform: https://chameleon-honeypot.vercel.app/

- **Placement:** 3rd Place
- **Team:** [Poushali Deb Purkayastha](https://cerebralvalley.ai/u/poushali), [Yukta Vajpayee](https://cerebralvalley.ai/u/yukta30), [Karthik Pragada](https://cerebralvalley.ai/u/karthikpragada06), [Garvit Sharma](https://cerebralvalley.ai/u/garvitsharma)
- **GitHub:** https://github.com/GarvitSharmaDeveloper/chameleon-honeypot
- **Demo video:** https://youtu.be/eV_S3T0LHhw
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=24

### 25. Jimmy 2.0

Jimmy 2.0 is a system that utilizes the YOLO vision model and dataset to identify objects it does NOT know, and then through a pipeline of Claude/GPT models and Browserbase research determines what that object likely is then embeds and stores that knowledge in a redis vector index for future reference. Utilizing weighted confidence metrics, Jimmy 2.0 is learning more about the world around it every time it goes for a walk.

Utilizing Daily - users can interact with Jimmy to learn more about what it has learned!

- **Team:** [Blake Ledden](https://cerebralvalley.ai/u/eklabbackwards), [Armin Foroughi](https://cerebralvalley.ai/u/arms)
- **GitHub:** https://github.com/arminforoughi/k1_coreweave
- **Demo video:** https://youtu.be/VoBni8OpK8s
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=25

### 26. Tryhard

Vapor - a platform where you can fine tune models for a specific usecase for browserbase automations, which will save cost over time

- **Team:** [Prajwal Umesh Nagalakshmi](https://cerebralvalley.ai/u/prajwalun)
- **GitHub:** https://github.com/prajwalun/vapor
- **Demo video:** https://youtu.be/0dKQTg_RlsQ
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=26

### 27. ZeroMe

Zero Me is a voice-controlled personal assistant with a cute floating blob UI that can manage your todos, calendar, documents, and emails via Notion and Resend APIs.
Self-Improving: The Personality Enhancer agent observes conversations, stores user preferences in Redis memory, tracks analytics to WandB, and can dynamically modify agent parameters (temperature, model) based on patterns like low task success rate or too many clarifying questions.

- **Team:** [Sanjay Sai](https://cerebralvalley.ai/u/sanjay_sai)
- **GitHub:** https://github.com/REDDITARUN/zero_me
- **Demo video:** https://www.loom.com/share/81e13e83b9c24228abcfc1985bd290f6
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=27

### 28. Catan Bias

Full browser Catan game built with React, TypeScript, Tailwind, and Supabase. Includes hex board, roads, settlements, cities, dev cards, trading, robber, longest road, and victory points. AI uses heuristic placement/build logic with optional LLMs (Gemini, Anthropic, MiniMax, OpenAI) for chat and trades. Python self-play (python/train.py) tunes weights and logs to wandb, exporting wandb-summary.json, which the frontend loads so LLMs and browser AI share the same training context. Marimo provides a dashboard to run games and track win rate and rounds.

- **Team:** [Petros Hong](https://cerebralvalley.ai/u/petroshong)
- **GitHub:** https://github.com/petroshong/catanlearn
- **Demo video:** https://youtu.be/n1E8BfAfQkM?si=5yszReCDLYbkWF1j
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=28

### 29. Protect The Weaves

Protect My Weaver is an AI agent observatory and cost management platform built on Weights & Biases Weave. While Weave provides distributed   
  tracing, we add the production control layer: budget tracking, anomaly detection, cost forecasting, and AI-powered analysis.

- **Team:** [Steve Kuo](https://cerebralvalley.ai/u/stevedusty)
- **GitHub:** https://github.com/Steve-Dusty/protect-the-weaves
- **Demo video:** https://youtu.be/vgnG_BYSq7s
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=29

### 30. Joyee Chen solo-person team

Since I have some AI evals experience but never tried evaluating agents before, I took this hackathon time to get started with the entire agent evals experience! I used UK AISI Inspect (very common in AI safety/AI alignment fields) and this github folder was basically started during the hackathon, showing my explorations. PreliminaryAgentEvalWork.ipynb lists some questions I wanted to convert to agentic evals, and TryingOutInspectAgentsDay2.ipynb was research workflow where I tried to iteratively improve an agent's performance on AssistantBench.

Note: because some extraneous circumstances forced me to a narrower project idea than usual as well as less effectiveness this time, even with AI coding assistants, I do not wish to be considered for prizes and thus will not send demo video nor be interviewed by judges. The "video" submitted to the Demo Video field is one I found online of AISI, the creator of Inspect, giving a talk at a wandb event about Inspect -- great job!

- **Team:** [Joyee Chen](https://cerebralvalley.ai/u/JoyeeChen1)
- **GitHub:** https://github.com/JoyeeChen/WelfareBenchmarkingRepo/tree/main/agent_eval_prototyping
- **Demo video:** https://www.youtube.com/watch?v=J79pSSAENYc
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=30

### 31. no slop is bad shader

Turn image into GLSL fragment shader code through self-improving agentic loops. https://shader-38ceag7cg-jessies-projects-a62a485b.vercel.app

- **Placement:** 2nd Place
- **Team:** [Jessie Han](https://cerebralvalley.ai/u/jessiehanvana)
- **GitHub:** https://github.com/JessieJessJe/shader-shade
- **Demo video:** https://youtu.be/1Zk7EOqembw
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=31

### 32. Jordan

JORDAN is a self-learning autonomous technical sales agent that eliminates deal-killing delays by providing real-time technical validation during live calls. Using Daily and Pipecat for ultra-low latency voice and Browserbase for live web-research, Jordan answers complex technical objections on the spot with the conviction of a seasoned closer. The system utilizes CoreWeave’s GPU infrastructure and Redis for lightning-fast context recall, ensuring every interaction is fluid and data-driven. Through Weights & Biases, Jordan continuously learns from previous sales calls, analyzing each call to refine his strategy and technical accuracy. He doesn't just assist the sales rep; he uses high-performance infrastructure to turn "I'll get back to you" into a closed contract.

- **Team:** [Aditi Apoorva Raghuvara](https://cerebralvalley.ai/u/aditijr), [Hrishikesh Athreya](https://cerebralvalley.ai/u/hrishi), [Arnav Dewan](https://cerebralvalley.ai/u/arnavdewan), [Rohan Hareesh](https://cerebralvalley.ai/u/rohanh903)
- **GitHub:** https://github.com/Hrishikesh-Athreya/wnbHack
- **Demo video:** https://youtu.be/IoiLQgaanUU
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=32

### 33. Interest Lens

"Your web, ranked for you."

AI-powered Chrome extension that personalizes any webpage by highlighting the content most relevant to you.

- **Team:** [Manoj M](https://cerebralvalley.ai/u/ManojM), [Ananta Verma](https://cerebralvalley.ai/u/anantaverma), [Jeffrey Risberg](https://cerebralvalley.ai/u/JeffRisberg)
- **GitHub:** https://github.com/iammanoj/interestlens
- **Demo video:** https://drive.google.com/file/d/1qlkadcvj3ssycDLWEmpgObKC5oKh7Y_2/view?usp=sharing
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=33

### 34. WebScout

WebScout is a self-improving web scraping agent that learns from every success and failure. Traditional scrapers break silently when sites change and repeat the    
  same mistakes. WebScout solves this with a 6-stage learning loop: it searches a vector cache for patterns that worked on similar pages, reuses high-confidence      
  matches instantly, or extracts fresh data with GPT-4o and Stagehand. Failed extractions trigger four recovery strategies including autonomous browsing and          
  cross-model Gemini analysis. Every successful pattern is stored as a vector embedding for reuse, and every step is traced through Weave. The result is measurable   
  improvement: cache hits climb from 14% to 83%, speed improves 3.7x, and success rates reach 100%.

- **Team:** [Nihal Nihalani](https://cerebralvalley.ai/u/nihalnihalani), [Yahya Alhinai](https://cerebralvalley.ai/u/yhinai)
- **GitHub:** https://github.com/nihalnihalani/WebScout
- **Demo video:** https://youtu.be/jWPj-5ritdo
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=34

### 35. We've Been Through This -- scalable smart memory mcp

Problem
We lose hours and millions of tokens getting stuck on conceptually similar issues in Claude Code. It gets even worse in teams where knowledge is shared ad hoc.
Solution
A team-shared MCP that conceptualizes where you get stuck and remembers.
A manual workaround is to add lessons to AGENTS.md, create skills, or write custom prompts - but a self-learning system is the natural evolution. To do it properly, you need to study history, extract knowledge, store it for the whole team, and inject it into the prompt.
That is what we do: trace Claude with Weave, analyze, store in Redis, and inject back into Claude.
The memory is hybrid-smart: the AI figures out which schemas and templates to use.

- **Team:** [Kirill Igumenshchev](https://cerebralvalley.ai/u/kirilligum)
- **GitHub:** https://github.com/kirilligum/we-ve-been-through-this-weave-hackathon-260131
- **Demo video:** https://www.youtube.com/watch?v=Z7IsFoGBSHk
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=35

### 36. Ghost Browser

We built a self-improving AI memory backend that prevents “fresh start” chats by continuously capturing a user’s goals, preferences, progress, and browsing context, then distilling it into clean long-term memories and reusable “what worked” patterns. When the user starts a new chat (ChatGPT/Gemini/Claude) or the conversation stalls, the system generates a compact context snapshot file and attaches it so the next model can continue instantly—while filtering unsafe requests, deduplicating repeated attempts, and only learning from validated successful outcomes.

- **Team:** [MAHIMA ADVILKAR](https://cerebralvalley.ai/u/MahimaAdvilkar), [Ashutosh Kumawat](https://cerebralvalley.ai/u/ashu)
- **GitHub:** https://github.com/DunngenMaster/weaveit
- **Demo video:** https://youtu.be/7yCclyWmJ-A
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=36

### 37. OneShot

OneShot is an AI-powered development platform that takes a user's idea (like "build me a todo app") and automatically handles the entire software lifecycle—research, technical specs, code generation, deployment, and testing—all visualized in a sleek stacked-card UI with real-time progress tracking.

The magic is in a Goal-Driven Agent (GDA) auto-recovery process: when any step fails, the system automatically detects the discrepancy, diagnoses the root cause, generates a recovery plan, and fixes the issue—all without human intervention. Every recovery leads to learning that compounds in a shared knowledge base.

- **Team:** [Dustin D](https://cerebralvalley.ai/u/dtdannen2)
- **GitHub:** https://github.com/dtdannen/oneshot
- **Demo video:** https://www.youtube.com/watch?v=Z_waiUFoWxE
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=37

### 38. Fractal

Fractal is an infinite curiosity engine that fundamentally inverts the search paradigm: instead of finding answers, it helps you discover better questions. 

It powers an intellectual journey from your initial thought to new understanding, using a suite of AI agents to extract concepts, generate lines of inquiry, and weave collected insights into new ideas.

W&B Weave: The backbone of our self-improving architecture. Using Weave traces, an LLM Judge scores the quality of every output, and the system updates its weights to favor higher-performing prompts over time without human intervention.

W&B Inference: Used for our agentic logic, powering the LLM models that handle Concept Extraction, Question Generation, and Synthesis.

The "Self-Improving Loop" is the core innovation. It treats prompt engineering as an optimization problem, using Weave to close the loop between generation and evaluation, allowing the agent to "learn" which questioning strategies yield the most insightful results.

- **Team:** [Mert Gulsun](https://cerebralvalley.ai/u/setrf)
- **GitHub:** https://github.com/setrf/fractal
- **Demo video:** https://youtu.be/0knDXb2XByU
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=38

### 39. claw

coding/completing tasks anywhere

- **Team:** [sam sam](https://cerebralvalley.ai/u/gguf), [Cal Cu](https://cerebralvalley.ai/u/calcu)
- **GitHub:** https://github.com/mochiyaki/claw
- **Demo video:** https://youtu.be/AvFN-JBXV2Y
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=39

### 40. Someone at my-side (SAM)

Voice assistant for ADHD/Autism support using Gemini Live API with Weave observability, custom evaluation scorers, and self-improving memory system.

https://docs.google.com/presentation/d/13ah4uzLODRytRen37BgEl5X1Sr8-H23Y-VBhBuUantI/edit?usp=sharing

- **Team:** [Vaibhav Dixit](https://cerebralvalley.ai/u/vaibhavdixit02), [Emma Qiu](https://cerebralvalley.ai/u/emq)
- **GitHub:** https://github.com/LingMQ/sam2-voice
- **Demo video:** https://drive.google.com/file/d/18rLXuvz2yUokfkzn_KagS2LMeMX7Bmpw/view?usp=drive_link
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=45

### 41. Clarity

Clarity is a self-improving, constraint-aware agent system that transforms raw photos of real spaces into realistic renovation visualizations without violating physical reality. The system analyzes each image to identify plumbing, electrical, HVAC, structural elements, and immovable fixtures so room geometry and utilities are preserved. This makes the outputs reliable for real architectural, commercial, and renovation use cases rather than just visual inspiration.

- **Team:** [Brian Permut](https://cerebralvalley.ai/u/dsl)
- **GitHub:** https://github.com/HiNala/Continuity
- **Demo video:** https://youtu.be/p3-N12w6BLE
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=47

### 42. mothBot

Mothbot is a selfevolving agent which maintains a space ship. It can call tools in chain instea dof running diagnostic on hull, then oxygen system and then gravity check, it cna just triggern them all in a chain in one tool call. We use marimo to run the tool chain engine its kind of like notebook, and then agent can save these tool chains as skills with descriptoin. So next time when you want to run diagnostics on hull, oxygen and gravity systems you can just reuse this, this also allows agent to fix the tool chains as agent knows what each tool is.

We use weaves to run tracing , and then we run evals at thread levels to add skills and then we do eval with coding agent to update prompts. This creates long running evolving loop.

We use pipekit for voice and agent orchestration, marimo for tool chain engine and then weavs for tracing and evals.

- **Team:** [vivek shukla](https://cerebralvalley.ai/u/weebhek)
- **GitHub:** https://github.com/vivek100/mothBot
- **Demo video:** https://youtu.be/Bx997YR7Ov0
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=40

### 43. loopless

LoopLess is a self-improving browser automation agent. It uses BrowserBase for cloud browser sessions with live streaming and permanent recordings. Gemini powers the LLM planning and serves as an LLM-as-a-Judge for evaluating task completion. W&B Weave provides full observability with tracing, custom scorers, and the built-in Evaluation framework to analyze failures and feed learned rules back into prompts. Redis caches successful action sequences as "macros"—cold runs learn, warm runs replay with zero LLM calls for 30-50% faster execution. The self-improvement loop automatically detects patterns (loops, missed actions) and improves future runs. Includes 18 AGI Inc benchmark tasks with real-time UI monitoring.

- **Team:** [Max Xie](https://cerebralvalley.ai/u/maxxie114)
- **GitHub:** https://github.com/maxxie114/loopless
- **Demo video:** https://youtu.be/KKCXkOnlggY
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=41

### 44. Duck Hunt Openenv

OpenEnv for thr VLM to play Duck Hunt game, ready for GRPO training and eval

Part of research that called:
Hardware-Aware Horizon Minimization for VLM Game Agents
Training Vision-Language Models to Play Duck Hunt with Adaptive Latency Compensation

- **Team:** [David Mayboroda](https://cerebralvalley.ai/u/davidmayboroda)
- **GitHub:** https://github.com/dmayboroda/horizon_min
- **Demo video:** https://youtu.be/bV8oiVYLhgs
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=42

### 45. Nova

We don't want to automate decision making. Since AI can write code, everyone knows planning becomes the most important step. However, no one wants to spend the time to challenge their assumptions or think about edge cases. This results in additional code iterations that slows down progress.

We built a tool that does two things:
1. Implements a robust planning & implementation process before writing a line of code
2. Improves our planning & implementation automatically after a coding session is complete.

The goal is the reduce the amount of steering during building, and this comes from good planning, not better coding models.

- **Team:** [Eric M](https://cerebralvalley.ai/u/Eric_m_cv), [Shoe Shoemaker](https://cerebralvalley.ai/u/shoe), [Matt Ruiters](https://cerebralvalley.ai/u/ruiters)
- **GitHub:** https://github.com/mcr5fh/Nova
- **Demo video:** https://youtu.be/7zwUR-NkgY4
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=43

### 46. Modaic

Microcode is an RLM-powered terminal agent built with DSPy & Modaic. It features a self-reflection loop where one instance monitors another's failures during execution. Using TerminalBench, I run 89 parallel Modal sandboxes with Claude Opus 4.5. A script: reflect.py watches for failures, curates test cases/trajectories, then calls Microcode to identify failure modes and generate:  FEEDBACK.md. Microcode then rewrites its own DSPy signature based on this feedback and pushes to hub. On startup, it pulls the latest revision, retaining improvements. The self-improvement cycle: monitor failed trajectories → analyze patterns → iterate on prompts → push updates → auto-propagate. RLMs handle the long coding trajectories. Code: github.com/modaic-ai/microcode | Engine: modaic.dev/farouk1/nanocode/tree/prod

- **Team:** [Farouk Adeleke](https://cerebralvalley.ai/u/farouk)
- **GitHub:** https://github.com/Fadeleke57/microcode-terminal-bench
- **Demo video:** https://youtu.be/FTbWSFsRAqY?si=LEb0ELkne0O1xAA9
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=44

### 47. Philoagent

Philoagent is a reasoning agent that answers philosophical questions. Chain of Thought + File System as Context + Swarm Browserbase MCP + Weave Inference and Tracing.

- **Team:** [Shawnick Wang](https://cerebralvalley.ai/u/shawnick)
- **GitHub:** https://github.com/shawnjuqi/philoagent
- **Demo video:** https://youtu.be/IFMj8dmB3f8
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=46

### 48. SIMPSEN

SIMPSEN is an autonomous software engine that builds and improves products without human developers. Give it a mission, roadmap, and a CEO personality. The CEO agent plans features and prioritizes work, then spins up dev agents in Vercel Sandboxes to write code, pass type checks/tests, and open PRs. A testing bot powered by Browserbase catches issues and feeds them back into the loop. A real-time dashboard shows live stats and a knowledge graph of your product’s evolution.

- **Team:** [Erik S](https://cerebralvalley.ai/u/eriksherman), [Yigit Guney](https://cerebralvalley.ai/u/yigitguney)
- **GitHub:** https://github.com/cleatai/simpsen
- **Demo video:** https://youtu.be/JyoDXWBXh8U
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=48

### 49. Conversationalist

This project acts as a memory extension for real life. It listens to conversations, extracts important details about the people you meet, and remembers them for future interactions. When you encounter the same person again, it recalls who they are and what you discussed before, allowing conversations to continue naturally instead of starting from scratch. The system addresses a common human limitation—forgetting names, faces, and personal context—by turning fleeting conversations into structured, searchable memory, making social and professional interactions smoother and less awkward.

- **Team:** [Parth Maradia](https://cerebralvalley.ai/u/Parth02), [Khush Patel](https://cerebralvalley.ai/u/khush007)
- **GitHub:** https://github.com/kxusx/conversationalist
- **Demo video:** https://youtu.be/vg2vxVVnCzI
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=49

### 50. Dave Remy Solo

**DML (Deterministic Memory Layer)** is an MCP server and Claude skill that gives AI agents structured, auditable, event-driven memory. Instead of appending facts to a growing text blob, every memory operation—facts learned, constraints added, decisions made—is recorded as an immutable event with a sequence number. This enables capabilities impossible with traditional memory: deterministic replay to any point in time, full provenance chains showing how facts evolved, counterfactual analysis ("what if this constraint existed earlier?"), and automatic policy enforcement that blocks decisions violating learned constraints. Built during WeaveHacks 3, DML integrates with Weights & Biases Weave for observability, where each memory event maps directly to a tracing span. Unlike traditional memory approaches where state is overwritten and history lost, DML preserves every event—so you can trace from any Weave span back to the underlying memory operations that informed the agent's decisions.

- **Team:** [Dave Remy](https://cerebralvalley.ai/u/daveremy99)
- **GitHub:** https://github.com/daveremy/deterministic-memory-layer
- **Demo video:** https://asciinema.org/a/poksmLtZA784MzjQ
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=50

### 51. Aetheris V.O.

Cloud platforms like AWS and GCP are powerful, but their interfaces can be complex and unintuitive - especially for new users. Our project makes navigating these environments effortless.

We’ve built an AI guide that lives directly on top of cloud UIs and helps users through voice and visual annotations. Instead of searching through docs or guessing where to click, users can simply ask and follow along. Our tool turns confusing cloud interfaces into interactive, guided experiences.

- **Placement:** 1st Place
- **Team:** [Reeti Bandyopadhyay](https://cerebralvalley.ai/u/Reeti), [Amogh Maheshwari](https://cerebralvalley.ai/u/Momacky90), [Raquib Alam](https://cerebralvalley.ai/u/theraq17), [Spencer Liu](https://cerebralvalley.ai/u/Spencer-L1u), [Savir D](https://cerebralvalley.ai/u/savur)
- **GitHub:** https://github.com/savir2010/aetheris
- **Demo video:** https://youtu.be/tfouA9-LRs0
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=51

### 52. ClinXplain

ClinXplain is an advanced medical platform designed to automate clinical documentation, streamline patient management, and accelerate medical research using state-of-the-art AI agents.

- **Team:** [Yuvraj Gupta](https://cerebralvalley.ai/u/Yuvraj), [pramod thebe](https://cerebralvalley.ai/u/pthebe)
- **GitHub:** https://github.com/YuvrajGupta1808/ClinXplain
- **Demo video:** https://docs.google.com/videos/d/1IV-00G7IwglZpGiOGDkxXZBD7VJpI0Byf7rgXuB4qI0/edit?usp=sharing, https://youtu.be/1P4BgugRkYQ
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=52

### 53. FULLSEND

FULLSEND

An autonomous GTM agent that ships ideas continuously, builds its own tools, and gets smarter over time.

- **Team:** [Balazs Nemethi](https://cerebralvalley.ai/u/balazs_nemethi)
- **GitHub:** https://github.com/nembal/GTM_agent_auto_mode_on/
- **Demo video:** https://youtu.be/2oNcfrhQDZk
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=53

### 54. OpenClaw Trace

OpenClaw Trace is a recursive self-improvement pipeline for OpenClaw / Clawdbot session traces. We mine real session traces for grounded signals - errors, friction, missed opportunities - then cluster them, generate research briefs, and run experiments to ship verified fixes. As a consequence, every improvement is evidence-backed, measurable and shareable - contributing to the body of science.

I'm already using this daily to recursively self-improve HAL, my OpenClaw AI researcher and virtual assistant.

Work done during hackathon: https://github.com/Phantastic-AI/openclaw-trace/pull/1

- **Team:** [Aditya  Advani](https://cerebralvalley.ai/u/ninjaa)
- **GitHub:** https://github.com/Phantastic-AI/openclaw-trace
- **Demo video:** https://screen.studio/share/8OCBBBBE
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=54

### 55. AgentHQ

A browser-based terminal coding agent multiplexer; to run agents where you want and control them from whatever device your prefer (mobile, desktop, laptop).

USED: Weave and Redis

- **Team:** [Rick Lamers NL](https://cerebralvalley.ai/u/rlamersnvidia)
- **GitHub:** https://github.com/ricklamers/agenthq
- **Demo video:** https://www.loom.com/share/7b3d190cbc6445b1a40a9698a1deb0a0
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=55

### 56. ExtraRecursiveReflection

Self-improving agents via accumulated reflection. Each iteration: diagnose failure (µf) → minimal skill edit → repeat. 

The key: Recursive LMs operate over the FULL history of attempts, diagnoses, and edits—not stateless reflection. The oracle learns to teach by seeing what worked.

- **Team:** [Darin Kishore](https://cerebralvalley.ai/u/darin)
- **GitHub:** https://github.com/darinkishore/rlm_reflection
- **Demo video:** https://youtu.be/wtfvu1uj2zc
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=56

### 57. GEB

My project is called Haggler and it is a negotiation/refund agent which autonomously self-improves such that you can get refunds without having the pressure of negotiating yourself. It uses Redis to store the different negotiation tactics, Pipecat for the voice agent harnessing, GCP for the speech to speech model, and weave for the observability layer.

- **Team:** [Kian Kyars](https://cerebralvalley.ai/u/kian)
- **GitHub:** https://github.com/kiankyars/haggler
- **Demo video:** https://youtu.be/TE4BKvLXE-I
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=58

### 58. DAP Notes

An AI-powered mental health documentation and exploration platform for therapists and potential therapy clients.

  For Therapists (/dap):
  • Verbally describe therapy sessions and automatically generate structured DAP (Data, Assessment, Plan) notes
  • Insurance-ready documentation with proper clinical formatting
  • CLI and web interface for flexible workflows
  • Built-in self-evaluation for iterative improvement

  For Clients (/intake):
  • "Is Therapy Right for Me?" - An interactive guided intake experience
  • Explores whether therapy might be helpful through thoughtful questions
  • AI-generated reflections normalize ambivalence and model collaborative interaction
  • Personalized outputs include therapy recommendations, first session guides, and practical experiments

  Features:
  • Full observability with Weights & Biases Weave integration for LLM tracing
  • Built with Next.js 15, React 19, TypeScript, and AI SDK v6 in a Turbo monorepo

- **Team:** [Cesar Toscano](https://cerebralvalley.ai/u/ctoscano)
- **GitHub:** https://github.com/ctoscano/cw-hackathon-3
- **Demo video:** https://youtu.be/5fVGzTtnpJY
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=61

### 59. Weave Got This

An AI-powered research and mapping app that turns natural-language questions into live map updates. You ask about places, routes, or current events in plain language (typed or spoken); AI agents run the research—searching stored findings and the web, with sources—then a map agent drives the UI by adding layers, pins, and viewport changes to a MapLibre map. The interface (map and chat) updates in response to the agents, so you see current research visualized on the map as the workflow runs. The project supports self-healing agents and workflows: W&B Weave traces research and map runs, automated scorers evaluate quality, and improvement reports suggest persona and config updates so agents get better over time.

- **Team:** [Anna Saltveit](https://cerebralvalley.ai/u/BananaDuck)
- **GitHub:** https://github.com/asaltveit/map-searcher
- **Demo video:** https://drive.google.com/file/d/1IU8G0S1-YxZNmbd8DQp1mh3Fh4dlVUgu/view?usp=drive_link
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=57

### 60. Atlas

A B2B internal platform for building, observing, and improving autonomous agents.
It provides trace-level visibility into agent runs, prompt and playbook management, and performance dashboards, enabling teams to debug behavior, enforce reliability, and iteratively improve agent decision-making in production.

- **Team:** [Kaiser Khoo](https://cerebralvalley.ai/u/silversky)
- **GitHub:** https://github.com/iceraider88/hack
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=59

### 61. Fotio

a self-reliant multi-agent system that fuses satellite signals and real-time phone calls to continuously evolve optimal wildfire prediction and response strategies using multi-agent evaluation and memory-based reinforcement learning.

we posted our demo on X and tagged weights and biases for the demo track: https://x.com/Ayaan_Gazali/status/2018081368820040168?s=20

- **Team:** [Ayaan Gazali](https://cerebralvalley.ai/u/ayaangazali)
- **GitHub:** https://github.com/aparekh02/fotia
- **Demo video:** https://youtu.be/aTGjUzFOXhI
- **Project:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery?project=60

---

Markdown version of https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt