I build and evaluate LLM systems, and I have spent eight years writing numerical simulation software for physical systems.

Most of what I do now is about getting models to do reliable work with tools. I build agents that operate external software, define the tool surface they are given, and improve them against test data rather than impressions. Most recently I worked at SPAR on inoculation prompting, a technique for suppressing unwanted behaviour in fine-tuned models. We found the suppressed behaviour is often still recoverable, mapped where it leaks back, and tested mitigations. The report is here and a paper is on the way.

Before that I built simulation software for novel simulations: two solvers from scratch in Python, C++ and Fortran, parallelised for cluster runs, and coarse-grained molecular dynamics models built inside a modified LAMMPS. I did a PhD in theoretical chemistry at Cambridge and a Marie Skłodowska-Curie postdoctoral fellowship at IST Austria, and taught statistical mechanics, thermodynamics and quantum mechanics at Cambridge for four years.

There is more on what I work on in the research section, and the full record in my CV.

Outside work I have spent a decade founding and running competitions and seminars for science students in the UK, Czechia, and Austria. Some of that is in activities.