Projects

Token Activation Maps for artwork description: a heatmap for every word a multimodal LLM generates, isolating the visual evidence that word draws on. Code and results for the ICPR 2026 paper.

• Interpretability
• MLLM
• Explainability

arXivCodeProject page

A multimodal agent that answers questions about paintings by retrieving evidence from Wikipedia and reasoning over it in context. Includes WikiFragments, a dataset of multimodal fragments mined at scale.

• Multimodal
• RAG
• Retrieval
• PyTorch

arXivCode

A Wikipedia-scale multimodal dataset — over 5M image–text pairs across 42M examples — built to give retrieval models something to ground on. Extracted with a modified WikiExtractor that keeps a page's images alongside its paragraphs. Released with ArtSeek.

• Dataset
• Multimodal
• Retrieval

arXivDataset

Text-guided multi-mask inpainting for artworks: a multimodal LLM proposes what belongs in each missing region, and a diffusion model paints it. Published at WACV 2025.

• Diffusion
• MLLM
• Inpainting

arXivCodeProject page

Few-shot semantic segmentation that takes points, boxes or masks as visual prompts and handles many classes at once. Built with the CILab team, published at ECAI 2025.

• Segmentation
• Few-shot
• PyTorch

CodeProject page

A small web app for looking at your own CLIP embeddings: project an image collection into two dimensions and browse the space to see what the model considers close to what.

• CLIP
• Visualisation
• Vue

Code

My MSc thesis work: generating descriptions of paintings from a noisy corpus of LLM-written captions, using CLIPScore to weigh each example by quality.

• Captioning
• CLIP
• Vision-Language

PaperCode

Worked solutions to the exercises in Bishop's Pattern Recognition and Machine Learning, typeset in LaTeX. Written while going through the book properly rather than skimming it.

• Machine Learning
• LaTeX

Code

Merged work on projects I use. Most of it is in Porcupine, a Python editor in tkinter, where six patches went in around the editor's cut/copy/paste behaviour and file-tree ordering. Also one patch each to scikit-learn (parameter validation for make_classification) and PyKEEN.

• Open Source
• Python

Code

A from-scratch encoder–decoder captioning model — the project where I learned how attention actually works rather than reading about it.

• Deep Learning
• Python

Code

A Texas Hold'Em engine in Python, written for the pleasure of getting hand evaluation and betting rounds exactly right.

• Python
• Games

Code

A soccer championship simulator — fixtures, tables, and a season that plays itself out.

• Python
• Simulation

Code