Writing

I write when a repo needs its argument written down. Each of these came out of a build, and each one is about the part I got wrong first.

Published

• The eval that could not tell a fix from variance · I planned to report a clean before-and-after. A second baseline run ruined the story, and that turned out to be the useful part. 5 min · Sep 2026
• Policy belongs in the tool layer, not the prompt · A rule you need an agent to obey should be arithmetic in code, not a sentence in a prompt. The call that taught me the difference. 6 min · Aug 2026
• Making an LLM safe enough to stand next to dangerous work · A voice agent that coaches a technician through a hazardous task, and how it was made impossible to talk into skipping the step that keeps them safe. 7 min · Jun 2026

Coming

• Threads, not facts · A unit of memory for a companion that has to remember you between conversations, not just during one. Drawn from the memory architecture.
• Discovery to go-live in 90 days · What it takes to deploy into an operator's existing systems, and why the unglamorous integration is the whole job.