Machine Learning Engineer · Computer Vision & LLM Systems
I build computer-vision and LLM systems and take them to production. In two years of industry work I built a multi-agent assistant serving hundreds of requests a day, an OCR pipeline that reads handwritten medical forms at 2,000 images an hour, and face-recognition serving on Kubernetes. I recently completed an Erasmus Mundus MSc in Computational Colour and Spectral Imaging (admitted 1st of 198) and co-authored a paper that won the Best Student Paper Award at IS&T Electronic Imaging 2026.
News
Selected Work
Trains on real photographs of one scene under different lighting so the recovered albedo keeps its luminance when the light changes: 28–39% less luminance variation in matched ablations and 7–14% lower albedo error on unseen coloured-light renderings, with 18.5M trained parameters. The next stage extends it to colour.
Co-authored at DXOMARK. Adapted a vision-language quality model to score portrait noise and texture separately, reaching SRCC 0.91/0.93 against expert ratings.
A multi-agent assistant serving hundreds of requests a day, with the failure rate cut from about 15% to about 5%. A Qwen3-VL-8B pipeline that extracts 14 fields from handwritten Japanese forms at 2,000 images an hour.