rhoGPT

rhoGPT is an umbrella org for my experiments with LLMs. only kara is launched right now.

1. kara -- a consumer app focused on providing the most realistic human-LLM interaction launched
2. ??? -- ai agent dev/deployment/observability platform TBD
3. ??? -- consumer app #2 TBD

whoami

i'm rohan. currently i am the only person working on this project. find my website here: rohan.ga.

email me at rohanganapa@gmail.com.

ideas i'm exploring now

1. factoring MHA/GQA weights into TPA weights for efficient inference
2. generalizing MHA past TPA
3. thinking about internal reward policies for GPRO vs extrinsic ones (ie ones that are judged by comparing to answers from a dataset)
4. automating prompt engineering for ai agents.
5. solving social media using psuedo-evolutionary algorihtims on top of LLMs where fitness == engagement
6. when UIs are neccessary and when MCP + claude code is all you need
   • what about when claude code can create UIs to help the user?

sources/insp: https://x.com/eddybuild/status/1895968364046020608, https://arxiv.org/pdf/2502.09992, https://arxiv.org/abs/2501.06425, https://arxiv.org/abs/2501.12948