PatchMind

Inspiration

I wanted to learn more about Site Reliability, RAG, and Agentic Ai so building a project like this was something that ticked all of my boxes. On top of this I'm a big web developer and I'm tired of having out of date repos after unreasonably short amounts of time

What it does

PatchMind is like an autopilot for keeping your code up to date. It scans your repo, figures out what’s outdated, digs up migration tips from docs and forums, and uses AI to rewrite code where needed. Instead of spending hours upgrading one library and hoping for the best, you get smart suggestions and safe patches.

How I built it

I broke it down into three main agents: Planner Agent — looks at package.json, finds the newest safe versions, and hunts down migration notes.

Parser Agent — maps out where dependencies show up across the repo so we know what’s at risk.

Codemod Agent — uses RAG to pull in the right docs and examples, then rewrites code when APIs change.

I leaned on Google Cloud’s agent-to-agent setup, Pinecone for embeddings, and built a simple Next.js UI. I also tried to bring an SRE mindset making sure indexing and crawling wouldn’t crash on big repos, adding retries and good logging from the start.

Challenges we ran into

Working with mulitple ai agents without losing data Handling huge codebases Being able to crawl many different documentation websites Writing reliable code with the A.I

Accomplishments that I'm proud of

Built a solid RAG pipeline that actually finds the right migration info when you need it. Learned a lot about SRE — making things resilient and observable instead of crossing my fingers. Created a dev tool that could save people hours of boring maintenance work.

What I learned

Multi-agent systems need clear communication and shared context RAG is good when done right and awful when done wrong.

What's next for PatchMind

Scaled up to be able to handle more larger repositories Improved vector retrieval for even higher code gen accuracy More A2A communication for more efficient workflows