Articles
Notes on AI evaluation, governance and quality engineering
Practical and specific, grounded in real engagements — not trend-chasing.
What "defensible" actually means in an AI risk classification
Most risk classifications collapse under one direct question: why this tier and not the other one. The difference between a classification that survives scrutiny and one that doesn't.
Ground-truth datasets are the part everyone skips
You can't evaluate a model against a standard nobody wrote down. A practical walkthrough of building a ground-truth set that actually holds up.
The EU AI Act's four risk tiers, without the legal drafting
Unacceptable, high, limited, minimal — what actually separates them, and the obligations that follow each one.
Why "it passed the demo" isn't a QA gate
Twenty years of test strategy taught one lesson that applies directly to AI: the happy path was never the hard part.
Why most AI governance policies are unenforceable
A policy document isn't a governance process. The gap between having a rule written down and having a way to check anyone followed it.
Benchmarking LLMs without cherry-picking the demo
A model that looks great on three curated prompts and a model that's actually reliable are not the same thing. What a fair benchmark set requires.