RSI to Karti Small.
Karti's PodMan entry placed second overall among more than 95 teams at the 2026 AI Engineer World's Fair hackathon and won the LiveKit and DigitalOcean tracks. The next step turned that recursive self-improvement result into a repeatable model-training loop.
Collect → Verify → Train → Evaluate → Improve
The first cycle used 512 verified rows and 64 optimizer steps against a pinned SmolLM3-3B foundation. On 26 August 2026, Karti-Small-RSI-3B completed the full path from reviewed data through training and evaluation.
The public repository records the model card, open-base lineage, and reproducible recipe. Private training traces, evaluation material, and Karti checkpoints remain protected. The program is open source, not open weights.