What This Is

I build and ship AI in industries where being wrong is expensive: healthcare, public services, law. I also research how to tell whether these systems actually work. This site is where those two habits meet, in writing.

Most of what gets said about AI is prediction. It will transform everything, or it will collapse under its own weight, and the person making the call is rarely around when the future arrives to check. Prediction is cheap precisely because nothing can disprove it in time to matter. For anyone who has to make a real decision this quarter, I find it close to useless.

So I write about proof instead. Not proof in the mathematical sense, but the ordinary kind: what actually held up when a system met real users, real data, and a real deadline, somewhere a wrong answer costs a patient, a benefit, or a legal exposure. That is a much narrower thing to write about than the future of everything. It is also the only kind of writing I have found worth reading when the decision was mine.

Who this is for

The person who is accountable when an AI decision goes wrong. In practice that is usually two people at once: the technical owner who signs off on the system, and the senior who carries the outcome and does not write code. I try to write for both in the same piece, deep enough for the first and plain enough for the second, because most of the failures I have seen happen in the gap between them.

If you want a forecast about where AI is heading, this is the wrong site. If you are trying to decide what to adopt, how to govern it, and whom to believe, you are in the right place.

What you will find

Essays, roughly one a quarter, each built around a decision you actually face rather than a topic. What survives production versus what only demos well. What responsible use and the EU AI Act require in practice. How to read a vendor’s confidence against the evidence. They run long, because the honest version usually does.

Every piece carries at least one concrete thing you can check: a number, a named clause, a study, or a real example with the details removed. If a claim on this site cannot be checked, I have not earned the right to make it yet, and I try to hold myself to that even when it is slower.

What I will not put here

• Predictions, futurism, and hot takes. If no outcome could prove it wrong, it is not a claim, it is a mood. I write about that failure mode rather than adding to it.
• Opinion without evidence. I have plenty of opinions. The ones I cannot support stay off the page.
• Client or confidential specifics. When I draw on real work it is anonymized, and the public sources are cited openly.
• Engagement bait. I would rather be read by a few people making real decisions than seen by many.
• Anything personal. This is about the work.

The standard

That is the whole thing, really. The AI conversation rewards people for sounding certain, and the work rewards people for being right, and those two are related far more loosely than anyone selling you a roadmap would like you to believe. I would rather be slower and checkable than fast and unfalsifiable.

If that is the kind of thing you want to read, the essays are through the door. If a specific decision is what brought you here, there is a way to reach me, and I only take a few.