AI node provisioning

1Problem

A 120B-class model does not fit in a 32 GB GPU, but it fits in a 128 GB unified-memory machine. Making consumer laptops into headless, always-on model servers on a rolling-release Linux takes fiddly work on each machine. By hand, it isn't repeatable.

2What I built

One provisioning script: 3,052 lines, 28 ordered steps, hardware detected at run time, with dry-run and verify modes. An offline test harness stubs the system commands and paths, so bootloader and hibernate logic runs on a workstation before it reaches hardware.

Claude's commands on the nodes go through a logging wrapper, and each node forwards its system journal to central logging. The nodes are tailnet-first: model APIs answer only over the overlay, and SSH is key-only. Network segmentation is designed, not yet enforced.

3How AI was used

Claude Code ran the research passes with quoted sources, wrote the script and the harness, and ran three independent reviews: 28 findings confirmed by a verifier agent in the first, 20 in the second, and 17 found in the last (16 fixed and checked by mutation testing). I did the BIOS work and the installs, and I gave each deploy its yes.

4Results

333 of 333 offline tests pass (v2.5 build log, Oct 2026). Two nodes are live. Turning off Wi-Fi power saving cut latency on the 128 GB node from roughly 90-120 ms to 2-6 ms.

5Status and limits

The generic-PC profile has never run on real hardware. Hibernate tests on the second laptop are still open.

6What I learned