One architect, not an agency roster.

I work with engineering teams on the part of AI systems that is not the model: how context gets assembled, how retrieval behaves on queries nobody anticipated, how you know whether a change helped, and what the whole thing costs per request at scale.

The pattern I see most often is a team that built something genuinely impressive in a few weeks, demoed it to real enthusiasm, and then spent six months discovering that the distance between a good demo and a dependable product is mostly architecture. Retrieval that was fine on twenty test questions degrades on twenty thousand real ones. An agent that handled the expected path has no defined behaviour for the unexpected one. Nobody can say whether last week’s prompt change was an improvement, because nothing measures it.

None of that is a failure of engineering talent. It is a failure of having someone whose job is the system as a whole — the boundaries, the failure modes, the feedback loops — while everyone else is correctly focused on shipping features. That is the role I play, for as long as a team needs it and no longer.

My day work is AI architecture inside a security company, which has shaped how I look at these systems. Once a model handles untrusted input, prompt injection and data exfiltration through model context stop being theoretical and become ordinary design constraints — the same category of concern as authentication or input validation. Most teams reach that realisation later than they would like to.

I work independently and stay that way deliberately. No vendor partnerships, no reseller margins, no referral arrangements. When I tell you a particular model or database is the right call, there is no second reason behind it.