I build AI systems for regulated workflows — compliance, audit, quality assurance.
Places where "probably correct" isn't good enough.

Currently: reinforcement learning environments where the reward signal is a human auditor's approval. If you can define what "correct" means in compliance, you can train agents that actually get there.

Previously at Yactraq: three LLM features shipped to enterprise production.

• Conversational analytics engine — 2x demo requests after launch
• Automated quality scoring — ~90% reduction in manual review, onboarding from weeks to hours
• Real-time compliance assistant — primary driver of new enterprise revenue

I write about what I'm building and what breaks along the way.