I work on AI evaluation and reliability.

Explainable where-to-live decision support for England & Wales. A tested dbt + DuckDB engine turns nine open-data sources into five transparent 0–100 indicators across 7,264 neighbourhoods, every score shown beside the raw figure it came from, served through a public FastAPI and a Next.js site with ~7k programmatic area pages. 228 dbt data tests + 2 unit tests, a versioned cross-runtime scoring contract, Dagster-orchestrated refresh, published lineage docs.

Works out when to run flexible electricity loads — a wash, an EV charge — to cut cost and carbon from live UK grid data, and shows the working: an LP/MILP optimiser behind a typed FastAPI contract, a Next.js planner, and a scenario retro that re-scores a plan under alternative carbon curves. Product development stopped in July 2026 after a scan found the scheduling function already served by a statutory smart-tariff duty and three shipping products — recorded with its primary sources rather than quietly abandoned.

Deployable end-to-end PHM MLOps, not another leaderboard notebook: NASA C-MAPSS turbofan RUL and ESA spacecraft-telemetry anomaly detection carried through their real evaluation protocols, wrapped in a FastAPI serving API, an operator console, signed release evidence (model card, SBOM, provenance), drift monitoring, and 462 tests. The evaluation layer proved general enough to extract as telemeval, a standalone library on PyPI with a Zenodo DOI.

A region-aware morning-market web app and an audit-ready, evidence-backed brief engine in one. The radar surfaces a market clock, a FRED overnight-risk rail, most-read finance news with AI summaries, and a Taiwan ETF-vs-benchmark attribution tool; the brief engine generates company briefs from SEC EDGAR + FRED, attaching a stored source span to every claim it accepts, with a click-through evidence ledger. Four localised editions (Taiwan, Korea, UK, EU). The CI gate that once certified this at 100% turned out to be scoring itself; measured against independently labelled ground truth, 40% of accepted claims are genuinely supported by the span they cite.

A synthetic neobank whose backend events misbehave on purpose — late, duplicated, reversed, schema-evolving — generated against a known-truth manifest so a governed dbt warehouse can be checked rather than trusted. On top sit the responsible-growth consumers: experimentation (CUPED, SRM, difference-in-differences, synthetic control), a calibrated activation model, and a release-gate that weighs customer-outcome guardrails. Run once on BigQuery: 68 dbt models under 217 data tests and 400 pytest tests, with full-refresh and incremental matching exactly at all six governed interfaces.

Which of a UK small business's unpaid invoices could break its cash runway, when the risk appears, and what to do this week: from an invoice CSV it forecasts a 13-week runway, ranks invoices by expected cash at risk with a plain-English reason on every score, enriches customer risk with Companies House filings, and writes the week's action brief — behind tested multi-tenant RBAC, a STRIDE threat model, and a DPIA.

Helps a UK ecommerce team decide where next quarter's marketing budget actually drives incremental revenue — turning noisy multi-channel data into cautious, uncertainty-aware budget recommendations via marketing-mix modelling, a Bayesian posterior layer, lift-test calibration, and a constrained budget optimiser, with a Streamlit executive dashboard.

Answers one question well: when London's transport is disrupted, how much extra demand lands on the bikes, and where? A 41.4M-journey PySpark backfill unified across five drifting schema eras, a tested dbt star schema, a LightGBM station-level forecast, and a free always-on live layer refreshed daily by GitHub Actions into committed Parquet — no warehouse to keep alive. Headline: strike days run about 1.4× median demand, up to ~2.3× on the worst full-network strike day.

Explainable where-to-live decision support for England & Wales. A tested dbt + DuckDB engine turns nine open-data sources into five transparent 0–100 indicators across 7,264 neighbourhoods, every score shown beside the raw figure it came from, served through a public FastAPI and a Next.js site with ~7k programmatic area pages. 228 dbt data tests + 2 unit tests, a versioned cross-runtime scoring contract, Dagster-orchestrated refresh, published lineage docs.

Works out when to run flexible electricity loads — a wash, an EV charge — to cut cost and carbon from live UK grid data, and shows the working: an LP/MILP optimiser behind a typed FastAPI contract, a Next.js planner, and a scenario retro that re-scores a plan under alternative carbon curves. Product development stopped in July 2026 after a scan found the scheduling function already served by a statutory smart-tariff duty and three shipping products — recorded with its primary sources rather than quietly abandoned.

A synthetic neobank whose backend events misbehave on purpose — late, duplicated, reversed, schema-evolving — generated against a known-truth manifest so a governed dbt warehouse can be checked rather than trusted. On top sit the responsible-growth consumers: experimentation (CUPED, SRM, difference-in-differences, synthetic control), a calibrated activation model, and a release-gate that weighs customer-outcome guardrails. Run once on BigQuery: 68 dbt models under 217 data tests and 400 pytest tests, with full-refresh and incremental matching exactly at all six governed interfaces.

Which of a UK small business's unpaid invoices could break its cash runway, when the risk appears, and what to do this week: from an invoice CSV it forecasts a 13-week runway, ranks invoices by expected cash at risk with a plain-English reason on every score, enriches customer risk with Companies House filings, and writes the week's action brief — behind tested multi-tenant RBAC, a STRIDE threat model, and a DPIA.

Helps a UK ecommerce team decide where next quarter's marketing budget actually drives incremental revenue — turning noisy multi-channel data into cautious, uncertainty-aware budget recommendations via marketing-mix modelling, a Bayesian posterior layer, lift-test calibration, and a constrained budget optimiser, with a Streamlit executive dashboard.

Works out when to run flexible electricity loads — a wash, an EV charge — to cut cost and carbon from live UK grid data, and shows the working: an LP/MILP optimiser behind a typed FastAPI contract, a Next.js planner, and a scenario retro that re-scores a plan under alternative carbon curves. Product development stopped in July 2026 after a scan found the scheduling function already served by a statutory smart-tariff duty and three shipping products — recorded with its primary sources rather than quietly abandoned.

Deployable end-to-end PHM MLOps, not another leaderboard notebook: NASA C-MAPSS turbofan RUL and ESA spacecraft-telemetry anomaly detection carried through their real evaluation protocols, wrapped in a FastAPI serving API, an operator console, signed release evidence (model card, SBOM, provenance), drift monitoring, and 462 tests. The evaluation layer proved general enough to extract as telemeval, a standalone library on PyPI with a Zenodo DOI.

An installable release-gate for AI agents (pip install agent-release-gates): replay known incidents, apply policy-as-code gates, and produce ship / warn / block evidence — as a CLI that fails CI, a UK AISI Inspect eval, or a runner pointed at your own agent's traces. Its most useful result is a negative one: the project's own synthetic benchmark turned out to be circular by construction, and the same retriever scores about twenty points lower on 640 external public cases. The external number is the one reported.

Adversarial benchmark foundry for LLM safety (pip install redteam-foundry): 1,883 prompts from four pinned corpora across 12 evaluation cells, with bootstrap confidence intervals and per-run API cost. Attack success came out at 0–4% — a negative result, so a known-vulnerable control model was pushed through the identical pipeline and scored 80%, which is what makes the zero mean something.

A region-aware morning-market web app and an audit-ready, evidence-backed brief engine in one. The radar surfaces a market clock, a FRED overnight-risk rail, most-read finance news with AI summaries, and a Taiwan ETF-vs-benchmark attribution tool; the brief engine generates company briefs from SEC EDGAR + FRED, attaching a stored source span to every claim it accepts, with a click-through evidence ledger. Four localised editions (Taiwan, Korea, UK, EU). The CI gate that once certified this at 100% turned out to be scoring itself; measured against independently labelled ground truth, 40% of accepted claims are genuinely supported by the span they cite.

Which of a UK small business's unpaid invoices could break its cash runway, when the risk appears, and what to do this week: from an invoice CSV it forecasts a 13-week runway, ranks invoices by expected cash at risk with a plain-English reason on every score, enriches customer risk with Companies House filings, and writes the week's action brief — behind tested multi-tenant RBAC, a STRIDE threat model, and a DPIA.

Helps a UK ecommerce team decide where next quarter's marketing budget actually drives incremental revenue — turning noisy multi-channel data into cautious, uncertainty-aware budget recommendations via marketing-mix modelling, a Bayesian posterior layer, lift-test calibration, and a constrained budget optimiser, with a Streamlit executive dashboard.