Four years shipping AI that runs in production, plus the unglamorous half: evals, latency budgets, and owning the thing once it is live.