Writing

Essays and teardowns on agentic AI in regulated workflows. New pieces every ~2 weeks. Home · Now

The calibration > correctness inversion in regulated-AI evals · upcoming

Draft · Q3 2026

Every frontier model fails calibration before it fails correctness. Why AI PMs in KYC/AML/CDS should gate on calibration, not accuracy — with 50 open eval cases.

Seven boundary violations that kill regulated-AI pilots · published 2026-07-03

Draft · Q3 2026

The failure mode isn't the model getting it wrong. It's the model getting it right at the wrong step of the workflow.

Designing a HITL gate a regulator will accept · upcoming

Draft · Q3 2026

What "human-in-the-loop" actually means when the loop has to pass an OCC or FDA audit — and why most current designs won't.