Machine Learning Engineer at Stats Perform. I build production LLM agents and the evaluation systems that decide what ships.
Led the evaluation platform for a pre-release agent. Evaluations on 54 service changes informed model and provider decisions and caught regressions before release. A scoring stack of deterministic fact checks plus an LLM judge cut structured-output failures from 8.0% to 0 of 1,744.
Added a GraphQL retrieval path alongside the legacy feed workflow. In a controlled 177-case A/B test, run errors fell 83%, input tokens 46% and median latency 30%.
Built and operate the service behind match summaries in the Opta Arena football app: four live data feeds, multi-provider routing with fallbacks, circuit breakers and cost caps. Nine months in production at a 0.18% 5xx rate.
MSc dissertation, Edinburgh. A leakage audit falsified the quote-level hypothesis and led to a calibrated case-level model, released with an open dataset of 24K quotes across 125 federal cases.
Stroke-care access mapping across 84K U.S. census tracts, presented as a first-author abstract at the International Stroke Conference 2024. 3D variational autoencoder for neuroimaging data augmentation.