Blog
Notes on infrastructure, AI agents, and taking prototypes to platforms.
Documentation drift, caught by audit
My architecture docs said our production nodes were one size and one cloud; reality was different in all four details. The commit that fixed it is literally titled "correct stale cluster facts."
Rewriting CI/CD service-by-service, without a freeze
We moved Aztecscan off Skaffold onto GitHub Actions + Kustomize one service at a time — and the Kustomize footgun that nearly broke traffic silently is the part worth remembering.
What an AI agent needs before you let it near production
Across two projects I gave AI agents what they needed to be trusted near production: the codebase's conventions in writing, a governed data path instead of ad hoc access, and a human interlock in front of the irreversible operations.
Parallel-run, don't cut over
Four real migrations where I kept the old path running until the new one was proven — and why the artifacts still in the repo are the proof.