k0bi

k0bi is the AI music-generation platform I worked on at Ozen. My part was the model-integration layer of the live generation pipeline — the seam where a request turns into text, style, audio analysis and a finished track.

It runs across more than six models: Claude for text and style, Gemini as an LLM judge for quality control and as a fallback, Gemma, an embedding model, audio analysis with Essentia, Demucs and pyannote, and Suno for the final render. I also co-built the backend behind single-shot and best-of-N generation, retrieval over the corpus, and the feedback loop driven by what listeners actually engaged with.

The quality work was the interesting part: tuning generation against an automated judge across roughly 450 end-to-end runs, with about 220 re-evaluated before release.