Fifteen years building calibrated evidence pipelines in clinical genomics -- designing high-throughput experimental programmes, extracting stable structure from noisy high-dimensional data, and turning uncertain observations into auditable decisions with formal error control. Now applying the same measurement discipline to agentic AI risk.
The structural transfer is direct. Both domains involve detecting rare consequential events in large populations, controlling false alarms under continuous surveillance, separating real signal from artefact through matched controls and replication logic, and producing evidence transparent enough for external audit.
Key contributions include the discovery of RAD51D and RAD51C as ovarian cancer susceptibility genes (now in global clinical use), quantitative evidence frameworks for clinical variant classification within CanVIG-UK, and population-level models that informed NHS cancer triage policy during the pandemic.
