Ozen
Model-integration layer for a production music-generation pipeline · 6+ models · corpus scaled 51×.
- Built with
- Python, RAG, LLM-as-judge
- Year
- 2026
- Status
- Archived
Ozen is where I worked as an AI engineer from February to August 2026, on a live music-generation product. I built the model-integration layer: more than six models in one pipeline, retrieval over the corpus, and an LLM judge doing quality control before anything shipped.
The piece I'm proudest of is the corpus. It started at 35 verified entries and ended at 1,806 — 51 times larger — through a pipeline that did the work instead of people: parsing documents and subtitle files, cross-verifying every entry against two independent sources, embedding it, and scraping on a schedule without supervision.
I also co-built the backend behind it — a monolith and two services — powering single-shot and best-of-N generation and an engagement-driven feedback loop.