I taught myself to build software the way you learn a language: move somewhere it's spoken, refuse to leave. So far that's a multi-tenant AI marketing platform, a local-first assistant on a runtime I wrote by hand, tools that quote jobs and answer reviews, and an RF field kit. All built end-to-end. All run by real people.
Along the way I noticed what separates AI products people demo from AI products businesses trust. It isn't the model. It's everything around it: the evals, the retrieval, the traces, the red-team. That became the work.
How I work
Measure before claiming. No number on this site was invented. If it isn't measured yet, the page says so.
Direct the machine, own the output. AI writes a good share of my code. I read every line and can defend every decision cold. That's not a caveat. It's the job.
Enforce at the lowest layer. Row-level security instead of an org_id
filter. A validation contract instead of trust. A tree-grep test instead of a convention.
Build by hand first. I wrote an agent loop from scratch before touching a framework. Abstractions leak. I know what's underneath.
Right now
A year of public work, one discipline per phase: evals, retrieval, observability, safety, agents. Each phase ends in a measured artifact inside a real system. Receipts on the work page, essays on the writing page as the numbers land.