Mathematics / AI / Research
I'm a research scientist at Visa working in the Risk and Security AI Lab (formerly Featurespace). This site serves as a home for current work, selected projects, and a lighter archive of earlier mathematics material.
Lucca Tower
About
My primary research at Visa focuses on the detection of machine-generated text. I'm also interested in how we generate text with language models, how we can use language models to understand the statistics of text of unknown origin, and how we can align language models with generative models for other kinds of data. Finally, I'm interested in the more human side of machine learning research, particularly explainability and fairness, where there are real questions to be asked about what we want to measure and how.
Previously I was an academic mathematician for twenty years, most recently as a senior lecturer (associate professor) of pure mathematics at the University of Manchester, where I worked in dynamical systems, ergodic theory, fractal geometry and number theory.
I previously held positions at St Andrews and Utrecht, and completed my PhD at Warwick in March 2011 under the supervision of Mark Pollicott. My thesis is available here. My Google Scholar profile is available here.
ML Research
Recent work has focused on deepfake detection, statistical text analysis, and the mathematical structure of decoding in large language models.
ICLR 2026
With Featurespace colleagues. How can one best view where a text sits in the next token distribution of a language model? DMAP offers a solution. Applications include understanding the overconfidence of instruction-tuned language models, verifying generation parameters for machine-generated text, and gaining a better understanding of approaches to detecting machine-generated text.
OpenReviewAISTATS 2025
With S. Burrell and C. Cheverall. We design a detector for machine-generated text based on local normalization distortion, exploiting a mismatch between the definition of temperature sampling in language models and similar, more realistic definitions in ergodic theory and statistical physics.
PMLREMNLP Findings 2025
With S. Burrell. A theoretical account of deficiencies in popular language-model decoding strategies using ergodic-theoretic language and an analysis of the distortion introduced by local normalization.
ACL AnthologyMathematics
Mathematics publications and preprint links remain available on a dedicated archive page.
Open mathematics researchI lectured courses in Metric Spaces and Fractal Geometry and taught smaller classes across all areas of pure mathematics and statistics. I have mentored postdocs including Lawrence Lee, Leticia Pardo Simon, and Demi Allen, and supervised PhD students including Alexandros Batsis, Peej Ingarfeld, Alden Paige, and Milo Edwardes.
I was on the London Mathematical Society Early Career Research committee, shaping how our research funding can best serve junior scientists. I was Pure Maths Research Lead at the University of Manchester. I organised Mathsbombe, a national mathematics competition for schoolchildren, as well as taking part in Science X and organising numerous research meetings and conferences.