About Me
I'm a PhD student at the University of Bristol, working in the intersection of Computer Vision and Natural Language Processing. My research focuses on Compositional Generalisation in Vision-Language Models, specifically understanding how these models compose and interpret spatial relationships between objects in images.
I completed a 3-month internship developing a tool to help radiologists identify meaningful differences in reports, supporting the training of junior doctors.
Before starting my PhD, I completed my MEng in Engineering Mathematics at the University of Bristol. After graduating, I worked as a Software Engineer at Wilxite, where I built internal web applications for data management, invoicing, and other business operations.
When I'm not coding or writing papers, I'm probably out running, rock climbing, or catching up with friends at the pub. I'm always interested in collaborating on interesting projects and discussing the future of AI!
Recent Work
Spatial Compositional Reasoning and Negation in Vision-Language Models
Status: In Progress ā¢ Expected: 2025
I'm investigating how vision-language models like LLaVA handle negation, by comparing how their visual attention shifts when processing positive vs. negated versions of the same sentenceāand how this compares to human patterns.
Semantic Similarity in Radiology Reports via LLMs and NER
Status: Paper Accepted at AI Bio Workshop at ECAI 2025
We explore how large language models can help evaluate junior radiologists' reports by identifying meaningful differences from senior-edited versions. Our method, Llama-EntScore, combines LLaMA 3.1 with named-entity recognition to produce interpretable similarity scores, achieving 93% accuracy within Ā±1 of expert ratingsāoutperforming LLMs and NER alone.
Evaluating Compositional Generalisation in VLMs and Diffusion Models
Status: Paper accepted at IWCS 2025
We explore whether diffusion models can better handle compositional generalisationāunderstanding how objects, attributes, and relationships combine in imagesācompared to models like CLIP. While diffusion models do well matching objects with attributes, allāincluding CLIPāstruggle with spatial relationships like "left" and "right." This highlights ongoing challenges in teaching AI to understand how parts of a scene relate.
News & Updates
š Best Poster at UK AI 2025
June 2025Honored to receive the Best Poster Award at the UK AI 2025 Conference supported by HDR (National Institue for Health Data Research). I presented my work on using LLMs to evaluate radiology reports.
š Attended CVPR 2024 in Seattle
June 2024Fantastic experience at CVPR 2024! Attended inspiring talks on computer vision, networked with researchers, and presented a poster on our compositional generalisation project. The conference was incredibly valuable for staying up-to-date with the latest developments in the field.