AI engineer working on evaluation and technical AI governance, at UCL Institute of Education.

I test whether evaluation results hold up: hold the model and the data fixed, vary the defensible choices made before the evaluation ran, and measure how far the conclusion moves.

Previously 14 years of cross-national policy research informing OECD, World Bank and FCDO policy.

TRACE_EVIDENCE: engineering components from TRACE. Deterministic preprocessing, cosine-centroid inference, drift and novelty monitoring, fail-closed acquisition, hash-bound release manifests, 43 offline tests.

education-newsletter-automation: live production ML pipeline and curator dashboard for a weekly education-policy newsletter. Held-out weighted F1 0.705, real-world weighted F1 0.630 on 264 manually-labelled production articles, with the gap analysed rather than omitted. Model card, threat model, drift and fairness monitoring.

pisa-specification-sensitivity: how far does a conclusion about educational attainment move across modelling choices that are all individually defensible? A 48-specification sensitivity study of PISA 2022 UK mathematics outcomes.

newstracker-template: reusable template for building domain-specific newsletter and news-tracker automation pipelines.

Live at trace-dashboard.onrender.com. A cross-jurisdictional NLP pipeline over almost 10,000 public policy documents from 63 sources across England, Scotland and Ireland. Validation uses a 580-document gold set, one expert human coder against an LLM gate, Cohen's kappa 0.743. Drift monitoring, and specification sensitivity reported per claim.

Research repositories are private because they hold corpora and interview material that cannot be redistributed, or because the associated paper is still under review.

sst-llm-judge: LLM-as-judge reliability study. Object of measurement held fixed, measurement setup varied across a crossed design. Private, code released on publication.

sst-public-sector-specification-testing: specification testing applied to public-sector AI use. Private, code released on publication.

AI-Governance-Analysis-UK-ROI: comparative analysis of how AI governance obligations are specified across the UK, Ireland, the EU and international bodies. 432 policy documents coded for deontic strength (binding, soft-normative, silent, absent) across seven specification dimensions, following Abbott et al. on hard and soft law, alongside expert interviews in the UK, Ireland and at UNESCO. Analysis in progress. Private, third-party copyrighted sources and interview material naming individuals.

• Evaluation methodology: LLM-as-judge reliability, benchmark design, specification range testing
• NLP pipelines for policy discourse analysis (NMF, BERTopic, Jensen-Shannon divergence)
• Construct validity and gold-set design for deployed classifiers

📧 louise@louiseyorke.com · l.yorke@ucl.ac.uk (until Dec 2026) · louiseyorke.com · LinkedIn