Codevolution

Inspiration

The RLM paper from MIT, finding novel applications for it

What it does

Gives AI a 20 million context window

How we built it

We used Daytona's sandbox to leverage Python's REPL to implement the recursive llm calls and then we supercharged the implementation with parallel execution. Then we made a UI that is great to see the agents being called in real time using a node visualization pattern

Challenges we ran into

Time and finding the right/visually impactful use case. We wanted to build this for codebase application, to allow the agent to understand the full codebase and generate the documentation and auto update the docs for every update, and we surely will after the hackathon. But for now, we just made a simple demo of the 20M context window

Accomplishments that we're proud of

We turned science into a real product. We actually made the paper implementation better.

What we learned

I learned that the people around the product matters most than coding itself. Great ideas leads creativity. Creativity leads progress. AI already codes better than us, so we learn to leverage what we as humans have the best to give.

What's next for Codevolution

Implementing the self-learning docs for codebases and probably some docs/due diligence application for finance companies that would benefit from huge context windows

DEMO: https://screen.studio/share/yzDnpYAy?state=uploading

Built With

• daytona
• nextjs
• ultracontext
• vercel