AI governance · applied AI systems
I build AI systems and the governance that keeps them honest.
38 of them since July 2023. 25 are running in production right now, 2 of those in the care sector, where being wrong has a cost nobody can refund. I own the whole line: the model routing, the evals that decide whether it ships, and the gate that stops it when it drifts.
No email wall on the resume. It opens in the browser and prints to PDF.
Live in production
Not screenshots. Not a private repo you have to take my word on. Running products, served to real users today.
Governance in practice
Governance is not a policy document I wrote after the fact. It is the part of the build that tells me no. Here is what that has actually cost me, system by system.
agent-gate
A portable safety fence for coding agents. 53 tests, and one design rule: any edge it does not recognise halts the run instead of guessing at it. That costs throughput, and I keep paying it. I would rather explain a stopped run than an unrecoverable one.
Find Your Vote
Address in, ranked candidates out, drawn from 25+ public-record sources with a per-issue score breakdown a voter can actually read. Determinism-audited: same inputs, same ranking, every run. A ranking I cannot explain line by line does not ship.
TradeTEST.TRAINING
~3,400 bilingual items through a five-stage QA gate, resumable and cost-capped so a bad run is cheap to abandon. When the eval regresses the build does not go out, and I do not get a vote on it. That is the point of writing the gate before the feature.
Multi-Model QA Cascade
One item routed to Ollama, Groq and Codex in parallel. A non-LLM scoring layer picks the winner, then a person approves before anything touches SQL. It is slower than letting the model commit. It is also why nothing out of this pipeline has had to be rolled back.
Every demo on this site
Each demo tile runs on fictional sample data. Real products operate on separate, isolated data paths. Two of the production systems sit in the care sector, so handling PHI under HIPAA is a constraint from the first commit rather than a retrofit before launch.
The unflattering part
A governance page that only lists wins is marketing. These are the two misses I think are most worth knowing about before you hire me.
A vision estimation tool I built scores itself against labelled ground truth. For months that eval counted every “not confident, retake the photo” response as a 100% error. So the accuracy figure I was reporting to myself was worse than the system actually was, and worse in the specific direction that would have pushed me to “fix” a model that was correctly declining to guess. A governance layer that penalises appropriate abstention is not a governance layer, it is a bug with a dashboard. I pulled retakes out into their own rate and re-baselined the rest.
The first version of my portfolio hub had a scrolling marquee and a counter that ticked up on load. Neither number was connected to anything. It read as proof and was set dressing. I cut both and rebuilt the hub around what the systems actually do.
Next step
If you are standing up an AI review function, inheriting a model nobody wrote an eval for, or trying to get a care-sector system past its own risk committee, that is the conversation I want. The form routes straight to me.