What you can actually hire me for.
Four ways I usually come in. Most engagements start with the first one and grow from there — the shape matters less than whether the thing ends up in production.
Architecture for AI that has to ship
Multi-agent orchestration, agentic loops, RAG and hybrid retrieval, multimodal extraction — designed against the constraints you actually have: data residency, PII, audit, cost per request. I do the design and I stay through delivery, because an architecture nobody can build is a document, not a solution.
Best when: you have a use case that works in a notebook and stalls before production.
Evaluation frameworks
How you will know the system is good enough to go live, and how you will know when it stops being so. Golden datasets and regression suites, tool-call assertions that test the arguments a model produces rather than its prose, LLM-as-judge rubrics where they earn their place, and validation programmes run with the domain experts who sign the output.
Best when: nobody in the room can answer "how do we know it's right?"
Governance and deployment posture
Where the inference runs, what leaves the building, who is accountable for the answer, and what the EU AI Act means for this particular use case. Hosted frontier models with region-pinned inference for most workloads; self-hosted open-weight models on private infrastructure where regulation demands it. The point is to choose deliberately, not by default.
Best when: legal or compliance has stopped the project and nobody knows how to restart it.
Technical pre-sales and enablement
I have been the fixed technical counterpart alongside sales on every opportunity of a Data & AI business unit — discovery, scoping, proposals, architectures presented to executives. And I train: a hundred-plus talks, years of courses, two companies taken from no AI at all to AI in daily use.
Best when: you are selling AI work, or your team needs to start doing it without you.