Topics I work on.

A short tour of the threads that run through my published work — what I find interesting, what I keep returning to, and the older papers behind the more recent ones.

Retrieval and representation learning

My PhD at EdinburghNLP with Victor Lavrenko was on hash codes for large-scale image retrieval — the line of work that became Variable Bit Quantisation (ACL 2013), Neighbourhood-Preserving Quantisation (SIGIR 2013), Graph-Regularised Hashing (ECIR 2015), and a 2019 monograph on the field. Sparse Kernel Learning for Image Annotation (ICMR 2014) won Best Student Paper. The throughline is the same question I still find interesting: how do you build representations that let you find the right thing in a huge collection, fast.

Some of my recent essays — on the bits-over-random metric, on why some RAG queries can't be solved by vector search alone — are extensions of that thinking into the LLM era.

Image enhancement and small interpretable components

A second strand is photo-quality image enhancement. DeepLPF (CVPR 2020) introduced learnable parametric filters, CURL (ICPR 2020) introduced neural curve layers, and SIDGAN (ECCV 2020) introduced a synthetic-data pipeline for low-light video. The shared idea: small, interpretable modules that get most of the benefit of much heavier networks.

Applied AI in regulated environments

For the last several years I have been part of JPMorgan's AI Research group, working on the deployment of generative AI in one of the most regulated computing environments out there. Published work from that period — SpamT5 (FinLLM @ IJCAI 2023), CodeQUEST (ISSREW 2025), API-Miner, Senatus, Ledgit, DeepClean — is mostly about making large models do real, audited work: code-quality evaluation, anomaly detection, federated training on sensitive data, machine unlearning. The patent portfolio is the residue of that effort.

How I tend to work

• Bias toward problems that ship. I have spent enough time near production systems to mistrust research that only works on a held-out test set. I prefer benchmarks that resemble the deployment environment.
• Small, interpretable components. DeepLPF and CURL are 600-line modules, not 100M-parameter monoliths. Many of the hashing papers replace a heavy lookup with a few learned bits. Where possible I'd rather make the inductive bias right than throw scale at it.
• Writing as thinking. I publish technical essays on Towards Data Science and Medium because the act of writing for practitioners surfaces gaps in my own understanding.

Full list of papers and patents on the home page.