"Values in UN General Debate Speeches" asks whose values national leaders assert on the UN podium. It classifies roughly 25,700 sentences of political speech from 61 countries against ten World Values Survey dimensions, then measures alignment with the surveyed values of each country's own public and with the Secretary-General's value assertions. Machine learning and NLP do the classification; eight years inside diplomacy supplied the operationalisation of the norm life cycle, norm clusters and soft power. Awarded a distinction. A methods-led journal article, with the error-propagation model as a second contribution, is in preparation.
On the Values for Cohesion project at the University of Birmingham, I contribute the quantitative measurement angle: translating a cultural intervention into measurable outcomes, defining what to measure, how to detect change, and what would count as evidence that it worked, then writing the results up for non-specialist audiences. The project is funded by the UK Ministry of Housing, Communities and Local Government through Birmingham City Council.
A running thread: whether published causal claims survive scrutiny. At LSE I replicated and stress-tested a published instrumental-variable study, reconstructing transformations the authors did not report and showing instrument relevance weakening materially under them. That exercise grew into SILICA-Bench, which turns the same audit into a measurable benchmark for frontier models.