From an RSI hackathon win to our first trained model.

The 2026 AI Engineer World’s Fair Hackathon centered on recursive self-improvement. Our team built PodMan, a voice teammate that watches a shared engineering room, flags duplicate work and merge collisions, and returns useful context to the team. It placed second overall out of more than 95 teams and won the LiveKit and DigitalOcean tracks.

Could the improvement loop teach a model we own?

Cloud models are extraordinary, but the Veronica and Prime Agent stack also needs a compact brain that can run locally, stay useful during an outage, call our tools precisely, and improve from reviewed work without exporting private household or operating context. That made a 3B model the right first target: small enough to iterate on, large enough to prove the whole system.

One completed checkpoint—and a repeatable path to the next one.

The first private checkpoint completed a BF16 LoRA supervised fine-tuning run from the pinned SmolLM3-3B foundation: 512 verified training rows, 64 optimizer steps, and a clean diagnostic evaluation. More important than a single score, the run established the complete path from curated evidence to a trained checkpoint and identified the next work: clarification coverage, strict response contracts, and tool-call structure.