Open tools for understanding the world and making institutions work better. Built with people I’ve learned a great deal from.
Codify
Making the law legible.
Every country should have a clear, citable, machine-readable account of its law. Codify is the work I’m doing to help make that possible.
The open-source core turns legal documents into structured text. The purpose is practical: help people understand their rights, help businesses navigate the rules, and give governments a clearer foundation for changing them.
It also draws on an older question that keeps bringing me back to Justinian: how does a society keep its laws understandable as complexity accumulates?
Codify: turning a statute book into structured, machine-readable law.
CivBench
A long game for AI.
Can an AI sustain a strategy, notice a rival’s progress and follow through on its own plans? Civilization VI gives those questions somewhere to play out.
I built the game interface and worked with collaborators to turn those experiments into a research benchmark. The diaries and tool calls let us inspect the decisions behind the outcome.
Accepted as a poster at NeurIPS 2026 Evaluations and Datasets Track
With Austin Tudor David Andrews, Jamie Heagerty, Harry Coppock, Jakob Nicolaus Foerster and Rui Ponte Costa.
A recorded Korea game. Bright flares show where the agent directed its attention as rival empires expanded. Explore the game diaries.
Learning inside government.
I led incubation — the early-stage projects — at the UK government’s Incubator for AI. The work was to find worthwhile problems with departments, test what AI could do, and help promising ideas become useful services.
How the Incubator worked
The system behind the projects.
I worked with colleagues on a practitioner’s playbook about how i.AI operated and what we learned. It connects problem selection, technical experiments and evaluation to the harder work of putting services into use across government.
Incubation was one part of that larger system: testing assumptions early, learning with users, and deciding which projects should progress, change direction or stop. The projects below grew out of that environment.
Alexander De Ville led the writing, with Alex Jones, me, Sharif Kazemi and Shahrukh Wani. The linked version is a discussion paper, open to feedback.
We built a way for civil servants to work with large language models and their own documents. It gave colleagues across the Cabinet Office room to experiment and develop informed opinions about AI.
I worked on it as an engineer and a first-time product manager. The service was retired as other tools became available; the experience shaped how I approached the projects that followed.
Lex makes UK legislation searchable through an API and tools that AI assistants can use. It brings legal sources into the same working environment as the questions people want to ask of them.
My work on Lex brought together engineering and management, applying lessons from Redbox to a more focused problem.
I helped build Parlex, a parliamentary intelligence service that helps civil servants find and understand what is being said in Parliament. Private offices use it to prepare ministerial briefings and debates, alongside parliamentary research by policy and bill teams.
Shared authentication helped the service spread across departments. Giving parliamentarians access to the same tools mattered to me too: better access to evidence should support both sides of the conversation.
Parliament MCP is part of the open-source foundation beneath Parlex. It connects AI tools to parliamentary data, including semantic search across Hansard and parliamentary questions, so others can build on the same work.
Our visit to Singapore’s GovTech sharpened my thinking about technical capability and shared infrastructure. Vikie Bew and I wrote about what we learned.
At the UK–ASEAN AI Innovation Summit in 2025, I joined a panel on AI in public services, exchanging experiences with colleagues working across the region.
I also helped lead the UK side of our exchange with France’s ALLiaNCE team. Farzana Chowdhury’s account captures the conversations, demonstrations and people behind it.
A tongue-in-cheek response to AI’s American spelling habits. It converts American English to British English in text and code files. A small contribution to keeping the civil service mildly less irritated.
