The dashboard is the summary. This is the version with the reasoning left in — who I am, what I take on, and how to reach me.
“I still find the work properly hard and properly enjoyable on most days, which I think is the only honest reason to keep doing anything for that long.”
I have been writing software for twenty seven years. I still find the work properly hard and properly enjoyable on most days, which I think is the only honest reason to keep doing anything for that long.
Outside the day job I like to train, and ideally die, in the gym. KBCoach came out of that: a coaching tool for kettlebell sport, built because the sport gave me a lot and I wanted to put something back into it. Schmooze is the other one, privacy-first AI on your own device. AI made both cheap enough to actually try, so now I am trying.
Most of my paid work is the unglamorous half of AI: making a model's output defensible when someone downstream has to sign off on it. Regulated financial services, mostly, where “the LLM said so” is not an answer an auditor accepts. I've spent 27 years shipping software that has to survive scrutiny, and the last few of them working out how to put language models inside those same control frameworks without pretending the hard parts aren't there.
RAG and summarisation pipelines where every answer carries its provenance — which document, which passage, why. Built for the case where a regulator, not a user, is the one reading the output.
The plumbing that keeps a model in scope: input/output guardrails, human-in-the-loop judgement gates, eval harnesses that catch regressions before they ship, plus prompt & context engineering and ADRs that record model provenance and the reasoning behind each call.
Introducing Claude Code and agentic development to teams that can't just point a coding agent at production. What it's allowed to touch, where the review gates sit, and how to keep an audit trail an engineering manager can actually stand behind.
Architecture review or hands-on build. I ship end to end on my own weekends — a paying multi-locale SaaS, a self-healing Sentry → Claude → PR pipeline, on-device Gemma — so I tend to spot the parts that break in production, not the parts that demo well.
Everything above is also available from the dashboard shell — try contact or consulting.