Designing controlled AI workflows for a regulated environment.

One controlled record of a firm's obligations, evidence and approvals, built privacy-first, with every finding traced to its source and every rule tested on history before release.

Agreed scope, fees, promised actions and approvals live in email. Drift shows up late, and the evidence is hard to produce when someone asks for it.

Lead product and AI engineer. I own the workflow design, the controls, the testing and the build.

• A pseudonymiser that codes people, organisations, places and contact details before any text leaves the machine.
• A project record built from the signed engagement agreement. Every field traces back to the words it came from, or is marked as not stated.
• Checks for out-of-scope requests, budget drift and fees against the agreement. Each finding quotes the sentence it came from and shows the arithmetic.
• An actions register that raises overdue and silent actions again, and a handover pack.
• Before releasing any risk flag, I replayed five candidates over a real 35-month email archive. One would have fired on about 1 in 6 outgoing messages, so it was redesigned before anyone saw it.

• Privacy first: names and contact details are coded before any text leaves the machine.
• Every finding is traceable to the sentence it came from.
• Rules are replayed on history before release.
• The AI layer is designed and next to build: Claude advises, the firm's own rules decide, and a person always approves.
• Nothing built so far sends client text to a model.

10,896

real emails replayed before a single flag was released

Source: rehearsal run in the commit history, Sep 2026

0 leaks

across 200 real messages and 1.6 million characters through the pseudonymiser

Source: test run in the commit history, Sep 2026

147

checks passing across 5 suites, 19 of them deliberate-break tests

Source: test run, 1 Oct 2026

Python, SQLite, Next.js, FastAPI, Supabase, Microsoft Graph (read-only)