I build AI features where measurement decides what ships.
Full-stack engineer since 2008. Co-founded and ran a real-time platform for a decade (100,000+ users, 99.9% uptime). Spent 2026 on full-time AI engineering R&D: RAG, agent runtimes, and LLM evaluation.
Four migration-shaped decisions across 573 of my own ADRs, and every one is a rewrite. Three never finished. The decade I spent upgrading Rails in place under live traffic is nowhere in that record, and the difference between the two is not taste.
I keep use cases as flat files: states, transitions, invariants, one file each. To find out whether the grammar travels, I ported one to electronic signature — envelopes, signers, webhook deliveries. The port broke a rule I had been leaning on.
I claimed an agent runtime with supervisor-owned gates. Here is the one wave that exists as evidence: a public bus directory, a merge sha, three counters, and the supervisor's acceptance written as a single line.
If one LLM call drafts and another judges, independent of what, exactly? My answer is weights lineage, enforced in code before any API call goes out.
A retrieval ladder for a Rails app, where every rung runs on a different model. How I picked which model sits on which rung, and how I keep that pick from quietly rotting.
