The pattern repeats in almost every conversation. A GenAI prototype exists, it works, and somebody has now promised it to the wider business. What follows is not a modelling problem — it is an infrastructure problem. Where does inference run. Who pays for the GPUs when nobody is using them. What happens to the data. Who signs it off.

That is the work below. Six areas, all of them things I have had to solve rather than read about, and all of them available separately — most engagements start with one of them, not the whole list.

Who this is not for

If you are looking for someone to fine-tune or train models, this is the wrong desk — I operate models, I do not build them. The same goes for pure data science work, and for teams that want a proof of concept to show a board next month with no intention of running it afterwards. Those projects can be done, they are just done better by somebody else.

Where this does fit: you already run software in production, you have a real workload in mind, and the constraint you keep hitting is infrastructure, cost or compliance rather than the model itself.

How an engagement usually runs

Almost always with a short assessment — two to three days spent on what you run today, where the AI workload has to sit, and what your compliance situation genuinely demands rather than what a vendor deck claims. The output is a written recommendation with an architecture sketch and a cost estimate, and it is yours to act on with or without me.

Implementation after that is scoped in blocks of a few weeks with something working at the end of each — a running cluster, a serving stack, a documented data flow. Not an open-ended retainer. Everything is handed over with the Helm charts, the Terraform and the write-up, because the point is that your team can run it once I am no longer around.