AI systems
AI that still works after the kickoff meeting.
I treat it like any other software: tests, monitoring, who can see what, and what happens when the model is wrong. Flashy demos are easy; boring production is the hard part.
Where teams get stuck
The prototype lands in a week. Then cost, reliability, and plugging into your systems eat the next three months. I spend my time there: the unglamorous bit that makes the thing usable.
What I actually build
Workflow automation
Steps with a human in the loop only where it still adds value.
Retrieval & grounding
Answers grounded in your documents and data, not generic web slop.
Evaluation & QA
Small test sets, regression checks, a bar for quality you can argue about.
Guardrails
Policies, redaction, audit trails, matched to how much risk you can carry.
FAQ
Is it all LLMs?
No. Sometimes rules, classical ML, or plain search beat a big model. I pick what fits the problem and your budget.
On-prem or locked-down cloud?
Yes. If your security or data residency rules require it, we design for that from the start.
Want to agree a first real milestone?
Say hello