I like building AI systems with confidence intervals of their accuracy. A system should refuse before it guesses: per-field confidence, provenance on every number, and a rule that routes the uncertain cases to a person.
I find the problem myself, build it, and run it in production. That work goes out under Itamih, bootstrapped and self-funded.
Clinic is the one you can't see here. A from-scratch practice-management system that replaced a 26-year-old on-premise setup at a dental hospital, in daily use by doctors, billing, and ops. Private repo, so the case study is the link instead.
Before this I built Fend, a B2B-pilots marketplace that went 0 to 1,000+ users. And before that, document extraction for mortgage origination at Fundmore. In SF.