Blog
Notes from building production AI.
Short, practical essays on the patterns behind the work — keeping human judgment where it matters, wiring AI into real systems safely, and measuring before anything ships.
Measured, or it doesn't ship
A demo tells you the system can work once. An evaluation tells you how often it works, on the cases you actually care about — which is the only thing worth deploying on.
AI wired into your systems, safely
AI that can talk about your systems but never touch them changes very little. The useful — and safe — version acts through a typed, permissioned, audited boundary.
Keep human judgment on the decision
The most reliable way to use AI on consequential work is not to remove the human — it is to move the human to the one step where being wrong is expensive.