mixtrainai (@mixtrainai) header
mixtrainai (@mixtrainai) profile picture

Post-training for everyone

Joined July 2025
Always a delight to work with teams pushing the new use cases on the platform! Great writeup by @abhinavvishwa on how they used GEPA based self evolving agent!
Shoutout to the @mixtrainai folks on building a genuinely nice compute abstraction and platform. As a follow-up to the self-evolving agent build (teapaw.org/abhinavvishwa/status/2…), where I was running the GEPA loop locally, ad-hoc, and on demand, I decided to platformize it a bit. There are a few established platforms out there to offload the runtime, data, and observability, while keeping the triggers local, e.g., modal, baseten, mixtrain, and more specialized and opinionated for my usecase, braintrust, raindrop, cekura, etc. I wanted to keep GEPA as application logic, so I decided to stick to the first bucket. Picked Mixtrain - partly to get some white-glove support from @dharmeshkakadia as needed. I was pleasantly surprised to find how easy to use (and close-loop with claude-code) it was with the provided Mixtrain skills (mixtrain.ai/docs/ai-tools/sk…). The entire thing took me close to 30 mins to repurpose from the previous version and deploy on the platform. The rendered eval side-by-side (as visible here) was v useful in verifying the improvements and picking the winner prompts, compared to my spreadsheet solution from earlier. A few other things I liked: - secrets are workspace-scoped and injected into the sandbox, so my provider key never touches the code or the upload. - the agent, the judge, and the reflector are each just a small python class with a typed run(), and they were each turned into a hosted, versioned endpoint from the signature. But, a few things that'd make it even more useful for agent workflows (feat-req): - A first-party LLM gateway. The external model catalog today is mostly media-gen (image/video/audio); for text agents you bring your own provider key. A hosted LLM proxy with key management + cost tracking would make this whole class of use-cases first-class. - Close the eval loop with human ratings. The side-by-side is great for eyeballing; letting me click a winner or drop a quick 1–5 inline would let me calibrate (or eventually replace) the LLM judge and export a golden set, instead of the view being read-only. net - for keeping the algorithm local while offloading the runtime, data, and observability, Mixtrain was a more than decent fit.
1
188