I work on LLM agents and causal reasoning. Before Oxford, I studied computer science at the Technical University of Munich with a minor in medicine — my other long-term interest is deep learning for healthcare.
Research
arXiv·Project page·Code·Model weights
Plan2Map is a 208-case benchmark: given only a UK planning document, a system must reconstruct the machine-readable geospatial boundary it describes. I built GeoPlanAgent, the tool-augmented agent system introduced in the paper — it reads the document, locates the site, registers the map against a national basemap, segments the drawn boundary with a fine-tuned SAM 3, and projects it to geographic coordinates, reaching 0.736 mean IoU where direct VLM baselines reach 0.108.
Open source
I'm the #4 all-time contributor to TransformerLens, the mechanistic interpretability library. Supported by a $5,000 Manifund grant, I helped build the TransformerLens 3.0 release, which grew the number of supported models from ~200 to ~9,000; my main contributions were the backwards-compatibility layer and weight-level Q/K/V splitting. I've also made smaller contributions to InterpretML.
Earlier
My BSc thesis, supervised by Tingting Zhu (University of Oxford), benchmarked models that synthesize ECG from wearable PPG signals and evaluated the synthesized ECGs on downstream sleep-apnea detection. I presented the results at the Institution of Engineering and Technology (recording).