Castings and findings in artificial intelligence.
A small workshop building tools, protocols, and systems for thinking with machines. We work on the harness around AI models — the structure, coordination, memory, and discipline that turn a capable model into useful work. We don't train foundation models; our work begins where theirs ends.
The practice is Maxime Roy, a software engineer who also builds instrument software — C#/.NET, C++/Qt, and Python for medical devices, machine vision, and digital imaging. The AI tooling here is his independent practice, started in 2022.
- Protocols — open agreements for how AI systems coordinate and interoperate.
- Tools — concrete, installable programs that implement them.
- Foundries — pre-assembled configurations of agents and tools for a kind of work.
- Published thinking — what we learn by building, written down, dated, and signed.
| Project | What it is |
|---|---|
| superharness | Multi-agent task coordination for Claude Code, Codex, Gemini & OpenCode — shared contract, lifecycle, handoff protocol, dashboard |
| obsidian-semantic-mcp | MCP server for Obsidian — semantic search over your vault (pgvector + embeddings) |
| token-diet | Unified installer and dashboard for four tools that each cut a different kind of AI session waste — RTK, tilth, Serena, ICM |
| pencil-sync | Bidirectional sync between Pencil.dev designs and frontend code |
| vidistiller | Self-hosted video transcription API — markdown output with timestamps |
More in the workshop, in development — Synod (multi-agent deliberation protocol), Morpheme (agent-session observability), Tessera (vault attachment memory): see artificemachine.com.