the find
− 0.6 mBeneath every project in this catalog runs the same small civilization of machines. The Tower — a dual-GPU workstation — does the heavy making: image diffusion, video generation, and the companion's voice fallback. The Brain — an M5 Max laptop — serves the large language models and the digital twin. The Keeper — an always-on Mac mini — keeps the graph, surviving reboots on its own daemons. The Vault — a NAS — holds twenty-eight years of photographs. All of it meshed on a private network, none of it listening on a public port.
cross-section
− 1.3 mThe publishing pattern is uniform and paranoid: public-facing services bind to loopback, and the brains talk to each other over the private mesh; a gate sits in front of anything reachable and fails closed; an outbound-only tunnel carries traffic to the world, so there are zero inbound ports anywhere and the kill switch is stopping one process. Watchdog scripts and scheduled tasks resurrect anything that dies on the Tower; boot-persistent daemons do the same on the Keeper. Deployment is a git push to a bare repo and a restart hook. A system-tray control panel starts and stops every service with per-service GPU pinning and a live VRAM readout. The spare, a laptop, keeps a warm voice server idling so speech synthesis never waits on a model load.
inscription
− 2.0 mField note, from the arbiter's spec: The newest institution is a GPU arbiter — a single authority over the big card's "heavy slot," allowing one heavy workload at a time under a system-wide memory guard. It is deliberately being built like infrastructure, not a script: step one of six is read-only — it observes, plans, and answers what-if questions, but enforces nothing until it has earned trust. And its constitution contains one absolute: the trading system's execution path is allowlist-protected and structurally untouchable. No render job, no companion, no experiment can ever preempt it. Some things you make fast; the important things you make impossible.