Quests

Most AI tooling makes existing work faster. The projects here change what the work is.

Iris stores architecture as data rather than diagrams, so a model can be queried, extended by agents, and rendered. This site is rendered from an Iris model, which is the simplest demonstration of the idea. Gantry runs staged, gated processes from a repository: the data is captured once and each gate's document is rendered from it. Braintrust builds agent personas from the published work of people you follow, so you can consult them when you need to.

Everything here is the practice I use every day, written down so other people can use it and so I can be held to it.

The ADR work extends Architecture Decision Records with the WH(Y) method, dependency tracking and governance metadata. It is the same discipline I introduced as standard practice at MSD, and it governs my own projects. Protocols sets out TDD, specs, feature branches and releases for agent-assisted delivery. Skills are the Claude skills I built and use.

Pipeline came first and has more stars than anything else I have made. It is retired, and says so here, because a method I have moved past is more useful as a marker than as a recommendation.

If the models that matter only run on infrastructure owned by a few companies, everyone else rents access. This theme is about the alternative.

The AI Blueprint is the oldest project here. In March 2024 it asked what a national AI strategy for New Zealand should look like, sixteen months before the government published one.

I should be clear about my part in the other two. Open Swarm is my specification for a mesh that runs small models on consumer hardware, and it has not been built. Sovereign Engine runs, and I did not write it: I set the direction, wrote the mesh spec, obtained the sponsorship and the hardware, and ran the build as a quest on Guild Hall while other people wrote the code.

The claim here is about getting other people's work pointed at something none of us would have built alone. It is not about the code.

Framing work as a quest rather than a goal changes who takes part and whether they finish.

Guild Hall is where Game Masters design quests for a community and members complete them: accept a quest, submit evidence, earn standing. Sovereign Engine was built as a quest on it. Campaign Mode gives you a party of six AI advisors, based on Dr Simon McCallum's Six Animals framework, to argue with you before your work goes in front of anyone who matters. Uncharted Quests is where it started: a career progression framework run as a role-playing game.

This theme is about how engagement is designed. The smaller things I have built for fun are under Side quests.

Questions I did not know the answer to, and the systems I built to find out.

Unsorry is the larger of the two: a swarm of agents writing Lean 4 proofs, checked by a compiler that rejects anything unproven. Git is the queue, the compiler is the judge, and no sorry survives a merge. Failed proofs are split into sub-goals and sent back out; verified ones join a library that makes the next proof easier. I lead it, and a community runs the swarm.

Machine Dream asks a narrower question: what happens if a model keeps thinking when nobody is prompting it? It answers with a GRASP loop, consolidation cycles and persistent memory, tested on a problem with a verifiable right answer.