Writing
Notes on shipping AI systems that survive contact with users.
Shorter than a handbook, longer than a tweet. The hard-won opinions that survive a year of running production AI agents: RAG, evals, safe-handoff, the unglamorous craft.
Everything here comes out of client engagements — retrieval systems that had to stop hallucinating, agents that had to hand off to humans without dropping the thread, eval suites that had to catch regressions before a stakeholder did. Each essay argues one engineering rule we now apply by default and shows the production incident that earned it. New entries land as new rules prove themselves in shipped systems.
-
Cite or refuse: the rule that keeps RAG honest
We treat hallucination as a failure mode of design, not of the model. Every answer either cites or says no, and the eval set is what enforces it.
-
We build the eval harness before the agent
Most agent demos work on a happy path and break on edge cases. We build the eval set first, so when something regresses you find out in CI, not in production.
-
Safe-handoff: the line between automation and triage
An agent that 'tries to be helpful' on a question it can't answer is worse than no agent at all. Safe-handoff is a first-class action, not a fallback.