I’m a senior research engineer in Cambridge, building production LLM systems for privacy-sensitive legal work at Luminance.
I lead development of Luminance’s proprietary legal LLM, Luna Crescent, which now replaces key third-party models in production contract analysis. I’ve taken it from data preparation and fine-tuning through evaluation, quantization, and serving, cutting LLM spend by roughly 50% and improving speed by about 30% without losing accuracy.
My research background is in dialogue understanding. At the University of Aberdeen and The Context Lab, I compared how LLMs and humans process dialogue context, and presented my first-author paper on it at EMNLP. More recently, I won a Meta Llama hackathon by training a medical reasoning model in under 24 hours.
I came to ML through engineering: I built and published a ferrofluid-core tilt sensor, studied electrical engineering in Budapest, and worked on Ericsson’s national-scale telephony server before finishing a first-class BSc in Computing Science in Aberdeen. Along the way I’ve built ML and software systems for healthcare, defence, and industry, including platforms used by the NHS.