Projects

Built with AI. Built properly.

Real products we've built, for clients and for ourselves. Each one shows a different side of how we deliver with AI: business analysis, clear specs, AI research, real tests and agents that check each other's work.

Invoice Pipeline

Invoices in. Checked, structured data out.

An AI pipeline that reads invoices, pulls out the data, lists exactly what it couldn't read, then checks the supplier is real before anyone pays.

See it work

How it was built

  • Current state mapped. We used business analysis techniques to describe how the client's invoice process works today.
  • Future state agreed. We worked through the future process with the client before building anything.
  • Brainstormed with AI. We explored approaches with AI to shape the design.
  • AI-refined prompts. We used AI to refine its own prompts and create the agent.
  • Low-code workflow. The agent was added to a low-code workflow that runs the process.
  • Tested on real data. We tested it by hand on the client's real documents.

Ingest · Categorise · Extract · Output · Verify

Zen Garden

Build a Japanese garden. Find out what it says about you.

Lay out a Japanese garden, rake the gravel and trim the plants. When you finish, you get a feng shui reading and a fortune about what your garden says about you.

Open Zen Garden

How it was built

  • Spec first. A written plan sets out the design, every screen, the scoring rules and how it will be tested. The agents build against it.
  • Research. Feng shui and zen garden principles were researched and turned into six scoring rules: five elements, bagua, chi flow, yin and yang, composition and care.
  • Test-driven. Every fix comes with a test that fails without it. Unit tests check the game logic. End-to-end tests use real mouse and touch input on desktop, tablet and phone.
  • Three-agent adversarial loop. A Builder agent writes the code. A Breaker agent attacks the app like an impatient user and must prove every bug. A Judge agent re-runs everything and returns pass or fail.
  • Watchdog. A supervisor stops stuck test runs, retries only infrastructure failures and reports flaky tests instead of hiding them.
  • Long-running agents. Agents work on their own in timed rounds: 40 minutes to build, 30 to attack and 30 to judge, for up to three rounds per milestone.
  • Designed. A Sims-style angled view, low-poly 3D models, generated ambient sound, and separate layouts for desktop and phone.

35 items · 6 grounds · 106 unit tests · desktop, tablet and phone

Thinking Tools

Every bias, razor and law worth knowing, with the story behind it.

A searchable guide to razors, laws, biases, fallacies and more. Each one comes with the story behind it, when to use it and how people misuse it.

Open Thinking Tools

How it was built

  • AI research. AI researched all 108 entries, using different prompting techniques to find the source, the real story and the common misuses of each idea.
  • Spec first. We set the goals and a strict format for every entry: what it is, where it came from, the story, when to reach for it, and what to watch out for. The AI completed every entry to that spec.
  • Built with an AI coding agent. The page, search and filters were built with an AI agent, using the site's own design system.
  • Designed. We designed the output into eleven colour-coded categories, with instant search across every word on a card and a Surprise Me button.

108 entries · 11 categories

Want to build like this?

We bring the same patterns to your projects: clear specs, real tests and AI agents that check each other's work.