Notes from production.
Long-form writing on what it actually takes to move enterprise AI from proof of concept to production. No trend pieces.
Why most pilots stall at security review — and the four artifacts that get a system through it.
Monitoring, evaluation, and audit logging — specified as engineering requirements, not policy prose.
Task-specific evals beat benchmarks. How to build a harness your auditors and engineers both trust.
Three-year TCO, switching costs, and the questions vendors hope you will not ask.
Permission-aware retrieval, grounded answers, and the unglamorous parts that actually decide quality.
Explicit authority, logged actions, and human escalation — what makes an agent acceptable to risk.
Instrument the outcome, attribute conservatively, and report cost per result — not cost per token.