Building now
Halley
Model risk and compliance teams use Halley to deeply understand their models, surfacing the interactions, subgroup effects, and conditional patterns that other tools miss.
AI interpretability research and applications.
I design algorithms that allow humans to understand machine learning models: what they learn from their training data, how they make predictions, and where they might fail in unexpected ways.
Previously I built and interpreted models on histopathology data for cancer diagnostics, and did my PhD on model-agnostic interpretability (methods that explain a model's decisions without needing to see inside it).
Since then I've worked on AI interpretability for: failure modes of large language models (one of those, a set of tokens that broke GPT-3, made the news); scientific discovery (if a model can predict something you can't, it knows something you don't); and high-stakes applications for financial institutions where understanding model behaviour is critical.
I'm based in San Francisco.
Now
Financial institutions use machine learning to make decisions with real consequences, and Halley allows them to understand and explain their models with unprecedented clarity.
Existing tools typically provide feature importances and score pre-specified subgroups – but that's not enough to really understand the patterns a model is doing. Halley explains the interactions a model depends on, the thresholds it leands, and the feature combinations where it systematically under- or over-predicts.
It also works on any kind of model – and this unlocks the ability to use more expressive and performant architectures that were previously out of reach due to their opacity.
Recent work
Interpretability
An Introduction to Exemplar Partitioning
A new method for finding interpretable structure in language models 1000x faster than sparse autoencoders.
Essay
Data-Driven Discovery vs Hypothesis-Driven Research
Hypothesis-driven analysis can only ever find what you already suspect is there – here is a case for starting with the data instead.
Case studies
New Publications in Immunology, Plant Biology and Meteorology
Three publications where models trained on scientific data found novel patterns, and our interpretability methods made them legible. These findings were discovered automatically by Disco and written up with the domain experts who gathered the data.
Essay
Scientific Discovery in the Age of AI
A survey of how AI is used in science today, from systematic literature review to autonomous agents.
Tool
Why Did ChatGPT Say That?
An efficient black-box method for attributing a language model's output back to the parts of the prompt that caused it, even if the model is hidden behind an API.
Discovery · Language models
SolidGoldMagikarp
The glitch-token work: we discovered a set of very strange tokens that made GPT-3 break in unexpected ways.
Media
Talks & interviews
Podcast · Latent Space
Automating Scientific Discovery
A conversation with Swyx about data-first science and seeing what models learn.
Podcast · The Human Podcast
How AI Will Unlock Key Scientific Discoveries
On personal motivation, how AI can (and can't) aid scientific discovery, and why interpretability is important.
Podcast · Tech Brew Ride Home
Leap Labs
A segment with Brian McCullough on interpretability, AI, and seeing patterns that humans miss.
Pitch · Deep Tech Momentum
Deep Tech Momentum
The pitch that won us the Novel AI track + many prizes at DTM.
Live · Presence
AI for Knowledge Discovery: Finding Patterns Humans Can't See
A live session on using interpretability to surface patterns that standard analysis misses.
Press
Publications
Exemplar Partitioning for Mechanistic InterpretabilityarXiv · 2026The decline in influenza antibody titers and modifiers of vaccine immunity from over ten years of serological datamedRxiv · 2026Automated discovery of patterns in T-cell receptor physicochemical signaturesbioRxiv · 2025Explaining surface-layer theory departures in marine flux profiles with data-driven discoveryarXiv · 2025Benchmarking the Discovery EnginearXiv · 2025Growth cost and transport-efficiency trade-offs define root system optimization in ArabidopsisbioRxiv · 2025Towards data-driven scientific discoveryAI4X · 2025Open problems in mechanistic interpretabilityTMLR · 2025The Discovery EngineLeap Press · 2025Model-agnostic interpretability (PhD thesis)St Andrews · 2024The SaTML'24 CNN interpretability competition: new innovations for concept-level interpretabilityarXiv · 2024Prototype generation: robust feature visualisation for data-independent interpretabilityarXiv · 2023Believe the HiPe: hierarchical perturbation for fast, robust and model-agnostic saliency mappingPattern Recognition · 2022Interpreting lymphocyte classification from Hoechst stained slides with deep learningCancers · 2022Weakly supervised learning and interpretability for endometrial whole slide image diagnosisExp. Biol. Med. · 2022Learning to describe: a new approach to computer-vision-based ancient coin analysisSci · 2020Understanding ancient coin imagesSpringer · 2019