selected builds
Work that can be opened, run, and challenged.
These projects connect AI research with delivery: concrete problems, inspectable implementation, automated tests, and explicit limitations. They are working artefacts, not claims of production readiness.
RSI Eval Lab
A runnable research prototype for auditing recursively self-improving AI lineages when the system may have incentives to change its evaluator or weaken oversight.
- Contribution
- Design, implementation, tests, and research agenda
- Evidence
- Deterministic auditor, synthetic traces, CI, optional Inspect task
- Limits
- Research harness; not containment or a deployment safety case
$ rsi-eval examples/safe_run.json
Run: safe-adaptive-lineage
Verdict: PASS
Held-out delta: +0.300
Efficiency: +0.0800 / 1k tokens
Next challenge level: 3
Findings: noneAgentMisalignment
A propensity benchmark for misaligned behaviour in LLM-based agents, developed with the Cambridge AI Safety Hub and contributed to the UK AI Security Institute's Inspect Evals repository.
- Contribution
- Benchmark co-development, experiments, analysis, and paper
- Evidence
- Realistic agent environments across five behavioural categories
- Finding
- System-prompted persona can shift behaviour as much as model choice
AI Engineering Journey
A dependency-free assessment that turns seven organisational capabilities into a transparent maturity score and phased transformation roadmap.
- Contribution
- Operating model, scoring method, CLI, example, and tests
- Evidence
- Validated input contract and deterministic roadmap generation
- Limits
- A diagnostic for prioritisation, not a maturity certification
Board AI Governance Toolkit
A lightweight toolkit that turns board-level AI oversight into explicit decisions, accountable owners, evidence, and review dates.
- Contribution
- Risk contract, validator, dashboard generator, and templates
- Evidence
- Deterministic scoring, overdue-review checks, and automated tests
- Limits
- Must be adapted to legal duties, risk appetite, and jurisdiction
1 decision required · 1 overdue review
Garmin Running Dashboard
An end-to-end Streamlit product that parses Garmin exports into training analysis while keeping personal health and location data out of the public repository.
- Contribution
- Product design, ETL, analysis, UI, synthetic demo, and tests
- Evidence
- 156-run GPS-free demo, desktop/mobile verification, and CI
- Privacy
- Uploads processed in memory; local exports and outputs ignored
working contribution
What I bring to a focused build
Turn an ambiguous institutional problem into a testable service hypothesis.
Build and integrate an agentic prototype with explicit data and permission boundaries.
Design task-specific evaluations, failure tests, and human escalation.
Connect technical choices to ownership, risk, and measurable outcomes.
Present the result clearly to technical, policy, operational, and executive stakeholders.
contact
For executive technology mandates, public-interest AI collaboration, board advisory, or the safe delivery of agentic systems - start a conversation.