Work · 02 of 03
Blackletter: UK AI governance, from the official record
A product we designed and built ourselves: a searchable, cited corpus of UK and EU AI regulation with a daily regulatory record, a weekly brief and a connector that plugs the whole thing into Claude, ChatGPT and Cursor.
Client
Blackletter (an Axis venture)
Sector
Legal and regulatory intelligence
Services
- Product strategy and research
- Full-stack build
- Data pipeline
- MCP connector
- Editorial newsletter
Stack
- Next.js 15
- Supabase
- Resend
- Model Context Protocol
- Three.js
- Vercel
The idea
There is no single UK AI law. The rules live across statutes, regulator guidance, codes and enforcement decisions, written for lawyers and scattered across a dozen official websites. The people who must comply, data protection officers, compliance leads and founders, cannot read all of it and cannot afford enterprise governance software.
Blackletter reads the official record of UK and EU AI regulation every day, so those people can ask one question and get a cited answer. It cites, or it abstains. It never guesses, because in this subject a confident wrong answer is worse than none.
What we built
- The Library: a corpus of official UK and EU sources only (ICO, GOV.UK, legislation.gov.uk, the FCA, CMA, Ofcom and MHRA, EUR-Lex and the EDPB, plus the NIST, OECD and Council of Europe frameworks), queryable in plain English, with provenance and licence recorded on every passage.
- The Terminal: a daily record of what changed, which regulator said it, and which deadline it moves.
- The Brief: a weekly newsletter typeset from the same record, with every claim traced to a primary source.
- The Connector: a remote MCP server and REST API so a compliance team can use Blackletter from inside Claude, ChatGPT or Cursor, the piece no UK competitor offers.
How it was made
We scoped the product with structured research before writing code, checked the niche against every UK competitor we could find, and ran a blind two-judge bake-off against the best free alternatives before committing to the build. The site and app share one Next.js codebase on Vercel; subscriptions run through Supabase and Resend from a verified sending domain. The ingestion model is deliberately conservative: an official-sources-only allowlist, credibility-tiered research, and human review before anything is published.
Why it matters for clients
Blackletter is the same discipline we bring to client work, applied to our own product: research first, a system of record rather than a demo, and a platform designed to keep running as the rules change. It is also a live example of shipping an AI product where accuracy is the whole point.