Safety-Critical Systems & AI Control Engineer Agent Safety · Runtime Governance · Assurance Engineering · Systems Analysis London, United Kingdom
I combine 18+ years of safety-critical railway signalling engineering with independent systems and software work focused on controlling how capable AI systems act, how their actions are evidenced, and how failures can be contained.
My central engineering question is simple: what must independently be true before a system is permitted to produce a consequential effect?
| Public work | What it demonstrates | Evidence status |
|---|---|---|
| Consequence-Bounded AI Control (CBAC) | A falsifiable systems-control hypothesis about containing the consequences of unsafe agent behaviour | Research hypothesis; external validation pending |
| Public adversarial evaluation — V1 | A reproducible 512-case synthetic input grid, explicit test contract, and three demonstrated limits | 27 locally passing unit tests; 0 in-grid expectation mismatches; not independent security validation |
| Synthetic AI control demonstration | Runnable Python example with mocked authority and completion checks | Public in-memory teaching example; no real authentication or effects |
| NEOM: How the Bindu Method Could Make NEOM Work | Infrastructure economics, capital staging, investment gates and downside modelling | Independent illustrative analysis; not an official forecast |
| Public evidence register | Clear separation of hypothesis, implementation, verification and unproven claims | Live public disclosure record |
Automated public example checks: GitHub Actions — synthetic control tests. Results are limited to the published mock example; see the evidence register for scope.
- AI control and agent safety: authorisation boundaries, tool-mediated actions and consequence containment.
- Assurance and verification: requirements, negative-path testing, provenance, reproducible evidence and regression discipline.
- Resilience and recovery: uncertainty handling, fail-safe behaviour, replay and recovery considerations.
- Systems engineering: constraint diagnosis, failure modes, dependency analysis and evidence-led design.
- Independent technical ownership: architecture decisions, AI-assisted software development, testing, critical review and documented limitations.
Working environment: Python, Bash, Git/GitHub, structured JSON contracts, test harnesses and safety-critical engineering practice.
Through Coal Tiger, I am independently developing research and engineering work on AI governance and control for autonomous and tool-using systems.
The underlying engineering repositories are intentionally private. This profile does not imply that the work is commercially deployed, independently certified or safe for production use. Only deliberately public and scope-limited artefacts are linked above.
My NEOM Bindu revitalisation case study examines how a large interdependent infrastructure programme might be made more economically resilient through productive-asset-first sequencing.
The illustrative model compares a $46bn industrial core with $24bn retained as discretionary later development, tests financial stress conditions and requires independent investment gates. Its claim is that NEOM remains theoretically possible under revised sequencing and verified demand, not that the full original vision is proven bankable.
The work prompted public discussion on LinkedIn, which is professional engagement rather than independent financial validation.
London Underground / Transport for London — Signalling engineering (2008–present) IRSE-licensed signalling professional working in a safety-critical operational discipline involving technical diagnosis, controlled processes, safe-state reasoning and engineering evidence.
Mercedes-Benz — Technical service, parts and warranty operations (2000–2007)
Problem → Constraint → Architecture → Implementation → Adversarial test → Independent evidence → Regression
The key principle is evidence before authority. A system should not be treated as correct simply because it declares success.
© 2026 Leon Maurice Browne. Public descriptions and illustrative code are shared for examination and discussion; no general open-source licence or grant of proprietary rights is implied.