What we do:
We build GitAuto, an AI Coding Agent that creates pull requests from issues on GitHub for software enginnering managers.
What I want you to focus on:
Beat the current #1 player, Honeycomb, on SWE-bench and make GitAuto the #1 AI coding agent.
Why this matters to you:
It’s gonna be an impressive achievement in AI coding space. Also, it’s incredibly challenging.
How to get there:
I’ll leave the strategy to you, but given our current situation, I think the steps would look like this:
- First, modify the code to work with SWE-bench.
- Check our current SWE-bench score.
- Analyze the mistakes and try to fix them with prompt engineering first.
- Simultaneously, implement "evaluation" to track if the prompt changes improve the score.
- Once score growth slows, experiment with other LLMs like o1 (Strawberry), Sonnet, or smaller models like Phi-3, or fine-tune GPT.
- If we run into GPU or cost limitations, estimate what’s needed, and we raise funds accordingly.
But I’m not an fine-tuning or mathmatical expert, so I’m looking for a co-founder who is excited to work on this. I’m not asking you to be the CTO or anything complicated. Equity is an equal split with 4-year vesting and 1-year cliff. Until we raise funds, there’s no salary.
Closing and the ask:
Interested? Here is the open-source repository: https://github.com/gitautoai/gitauto
Also I'm open to chat: https://calendly.com/gitauto/wes