In probabilistic systems, the eval set is the product spec. Director of Product, Lyft Business — the PM who owns the eval set, the failure-mode taxonomy, and the recovery UX. Field Kit on this site is the live demonstration: three Claude-powered tools with per-task model routing, Zod-validated streaming, and per-run TTFT + token + cost observability. Fifteen years scaling AI-native and platform products across Hims & Hers, Dropbox, Serve Robotics, Agtonomy, and Just Eat Takeaway.