Koma-Mind

Use videos from your favorite teachers and tutorials creators to generate new tutorials in their style.We use a self optimizing agent using GEPA prompt optimization techniques.

Comment

Inspiration

This paper : https://arxiv.org/abs/2507.19457 A genetic prompt optimizer that shows better performance over fine tuning.

What it does

We use GEPA as implemented in the DSPy framework to fine tune video extraction prompts.

How we built it

Google ADK, Google Gemini 2.5 Multimodal for video analysis., DSPy for the optimization framework. Language : Python LLMs : Gemini 2.5 and GPT5

Challenges we ran into

Video analysis is SLOW!! Also gemini tends to error out for complex prompts

Accomplishments that we're proud of

Prompts are actually getting better and extracting relevant info! All from a generic starting prompt.

What we learned

Lots .... don't work with video at hackathons :-)

What's next for Koma-Mind

This will be used as part of our Neurotech BCI EEG brainwave analysis toolset

Built With

Share this project:

Updates