Technical notes from real build work — measured results, honest costs, no growth-hacking.
Why I built Verimem — a memory layer with a notary at the door
There were already five popular memory layers for AI agents. I built a sixth anyway, because all five share the same missing piece: none of them checks whether a memory is true before storing it.
Read →2026-07-04The reviewer must not be the author's friend
Why I built an adversarial code-review gate where each critic runs as a fresh AI instance that shares as little context as possible with the one that wrote the code — and why 'ask the model to check its own work' is worse than useless.
Read →2026-07-04What months of measuring confabulation and hallucination in LLM agents taught me
Five lessons from benchmarking memory systems against hallucination — including the time my own metric lied to me by 4×, and the false-memory attack that fools every entailment judge.
Read →2026-07-04The Defender's Tax: a security argument for open AI
Frontier AI safety taxes the defender far more than the attacker. A structural critique — no conspiracy required — and why security specifically needs capable, open AI. Draft.
Read →