AI Governance · Agentic Delivery Assurance · Responsible AI
I build software with fleets of AI agents, and I govern them while they work — ownership boundaries, least-privilege isolation, human approval gates, and controls enforced in CI rather than written in a policy nobody reads.
Nine years running regulated delivery at Nordea, Santander and Fidelity International taught me what an auditor asks for. The last four months taught me what actually breaks when you hand autonomous agents write access to a real codebase.
Around 98% of organisations have deployed AI agents or plan to, and fewer than half have any inventory of them. No owner, no scoped permissions, no revocation path, no audit trail. Agent sprawl is now a board-level risk and most governance functions are still writing principles documents about model bias.
The harder question is operational: who approved what your agents changed, how do you know, and can you show me?
That question needs someone who has actually run agents in anger and knows what a regulator expects. I sit in that overlap.
📘 Written up in full: Governing Coding Agents — A Field Guide — seven controls, with an ISO/IEC 42001 mapping.
This is not theoretical. It is the operating model behind TrackGemz, running four autonomous agents against one production monorepo.
Named ownership. Four agents — foundation, libkeeper, srwriter, jrwriter —
each with a written scope and a documented ownership test for deciding which one owns a
given change. No agent acts outside its boundary.
Least-privilege isolation. Every agent runs in its own git worktree on its own branch. Blast radius is one branch, always.
Controls enforced mechanically, not documented. Architecture boundaries are asserted by
tests (test_architecture.py in every package) — engines cannot import each other, core
cannot import upward. The control fails the build, so it cannot quietly rot.
Human approval gates. Agents run headless; autopilot is off by default; a person merges or discards at Review. The worktree and the gate are the safety boundary, and that decision is written down, dated, and justified.
Provenance. PROVENANCE.md in each package pins exactly which upstream commit ported
logic came from. Lineage is recoverable.
A dated decisions log. Architectural stances recorded with the date and the reasoning
(D1 — No SvelteKit (2026-06-10), D2 — Backend stance, …), so a decision can be audited
later against what was known at the time.
Incident-driven standing laws. Rules earned by failure, each carrying the incident that produced it — "cover text never goes into the picture; a book was blocked for a day proving it." Corrective action, recorded, with cause.
That is an agent registry with least-privilege isolation, named ownership, lifecycle control and a human approval gate. Mapped onto ISO/IEC 42001 it is most of Clauses 6–10. I arrived at it because I needed it to stop losing work, which is the only reason controls ever actually get used.
3,213 commits across ten projects, 13 May – 9 September 2026. Solo, agent-assisted, under the controls above.
| Project | What it is | Evidence |
|---|---|---|
| TrackGemz | Commercial macOS DJ-library suite — shipped, sold, auto-updating | 976 commits · 721 PRs · 269 test files · 22 public releases |
| Selecta | The monorepo it grew out of | 1,316 commits · 116 test files |
| Hotelier | Multi-property hotel PMS with anti-theft controls | 115 commits · 338 tests |
| Puzzle Engines ⭐ | 7 deterministic generators that re-solve their own output before returning | public · 95 tests · contract-enforced conformance |
| hq | Second brain + agent mission control with human merge gates | 132 commits |
| Long-form content pipeline | Staged generation with critique and revision passes, cost-measured per run | 398 commits |
| Lazyfilm | Brief → shot list → AI clips → edit-ready folder | 110 commits |
Plus ~40 custom agent skills — I build the tooling, not just consume it.
Most repositories are private (commercial). Happy to walk through architecture, controls or decision logs on a call.
Nine years of delivery leadership in exactly the regulated environments where AI assurance is about to be mandatory:
- Cloud Heights Solution — Scrum Master, banking: Front/Middle Office, Risk, Finance, Regulatory
- Fidelity International — Delivery Manager, cloud-security programme scanning all cloud resources and IaC
- Into Film — Head of Product & Technology, rights-compliant streaming platform
- Nordea Bank · Santander — Scrum Master, contract
Certified Scrum Master · Certified SAFe Practitioner
I know how governance actually lands in a bank: who signs, what evidence satisfies an audit, and why most control frameworks fail at the point of delivery.
Building toward AI Assurance / Responsible AI leadership, and consulting on agentic governance for regulated organisations.
In progress: IAPP AIGP → ISO/IEC 42001 Lead Implementer → AI security (OWASP LLM Top 10, MITRE ATLAS, agent threat modelling).
Open to: AI governance and assurance roles, and consulting engagements on agentic AI controls. Remote or UK-based.