[ writing ]
Notes on building software you can check
How I build AI tools that have to prove what they say, and what I learned shipping code at work.
- October 17, 2026 · 4 min read
Why my AI agent isn't allowed to check its own work
In Casefile, an AI writes anti-money-laundering case notes, but plain code checks every number against the ledger before anyone sees them. Here is how it works, and the bug that got past it first.
- October 17, 2026 · 4 min read
Proving a root cause before reporting it
RootCause is an on-call agent that finds why a microservice broke, then re-queries the telemetry to prove it before it says anything. Why the loudest service is usually the wrong answer, and why "I don't know" is a feature.