Full talk · edited transcript

Doubling Medical Safety: Fine-Tuning Open LLMs for Women's Health Without Human Labels

Vladislav Nedosekin, Michael Shtelma, Andras Meczner16 Jun 2026

Speakers: Vladislav Nedosekin (Director of Engineering, Data and AI Platform, Flo Health), Michael Shtelma (Lead Product Specialist for GenAI, Databricks), Andras Meczner (Medical Director for Accuracy and Safety, Flo Health). Databricks.

The talk covers four parts: the business challenge, how results were evaluated (what "good" looks like), the fine-tuning approach, and the outcomes with next steps. According to the session description, Llama 3.3 70B was turned into a healthcare-compliant assistant for women's health that doubled safety compliance versus the previous iteration; instead of spending expert time on labeling, it went into designing LLM judges that scale (60 judges: 52 for medical safety, 8 for usefulness, with priority-weighted reward aggregation where P1 safety rules dominate P2 quality rules).

Continue reading — it's free

Google or a one-time email code. No password.