Pragmatic AI features, shipped into real products.
No moonshots. I help product teams integrate LLMs, embeddings and classical ML into software they already ship — with an eye on cost, latency, evaluation and the very real risks that come with putting generative models in front of customers.
Where AI actually earns its keep inside day-to-day products.
LLM features
Summarisation, classification, drafting and extraction — wired into your existing stack with proper fallbacks and evals.
RAG & knowledge search
Retrieval-augmented assistants over your docs, tickets or codebase. Honest about what embeddings can and can’t do.
Agentic workflows
Tool-using agents for internal ops — careful scoping, audit logs, human-in-the-loop where it matters.
Predictive & classical ML
When regression or a gradient-boosted tree is the right answer, we use that instead of a 70B model.
Evals & observability
Golden sets, regression tests, latency & cost dashboards. Knowing when a new model is actually better.
Strategy & advisory
Short engagements to pressure-test an AI roadmap, a vendor choice, or a due-diligence question.
A sceptic’s approach to a hyped field.
-
01
Problem first, model second
If the problem can be solved with a SQL query or a rule, it probably should be.
-
02
Measure before you believe
Every AI feature ships with an eval set. Vibes are not a quality bar.
-
03
Guardrails, not prayers
Structured outputs, rate limits, PII filtering, and clear failure modes.
-
04
Own your evals, not your weights
Use hosted models when sensible. Self-host when privacy, cost or latency demand it.