AI advice is worth nothing unless you run it yourself.
Clients ask me what AI means for their estate — GPUs, agents, data gravity, cost. I treat the answer as a professional discipline, not a hobby: a syllabus, a lab bench, and a habit of measuring instead of admiring.
Continuous learning, with a curriculum
"I follow AI news" is not a learning strategy. This is a working syllabus — each block has a deliverable, and each deliverable ends up in a real customer conversation.
- Foundations: transformer architecture, attention, tokenisation, training vs inference economics — explained without hand-waving.
- Serving & infrastructure: GPU sizing, quantisation, KV-cache behaviour, batching — the pre-sales math behind every AI TCO model.
- Retrieval & agents: embeddings, vector search, RAG pipelines, tool-use patterns — and where each quietly breaks in production.
- Evaluation: golden datasets, regression tests for prompts, hallucination budgets — because an unmeasured model is a liability.
- Governance: data residency, model provenance, EU AI Act categories — the questions procurement actually asks.
The lab bench
Two local rigs, one purpose: every claim I make to a customer has run on hardware I own first.
Home lab
Applied in pre-sales
“The model is the easy part. The estate around it is the sale.” Working principle no. 4
Five questions, honest answers
"Do we need GPUs?"
Usually later than the vendor says, and smaller than the slide claims. Inference economics first; silicon second.
"Build, buy, or rent?"
Depends on data gravity and residency, not fashion. The map of your data decides more than the benchmark table.
"Is it safe to ship?"
Not without evaluation harnesses and an owner. I show the failure modes before the feature demo.
"What about our team?"
AI projects fail organisationally, not technically. The operating model is designed in the PoC, not after it.
The learning curve is the career arc.
A decade of enterprise pre-sales, three technology houses, one consistent habit: prove it before you promise it.
