FOU//CAL

~/blog

10 transmissions on record.

CyberNov 21, 20266 min

Give your AI agent a forge, not your SSH keys

My agent commits code every day. The lazy way to wire that up is an SSH key — a credential with no scope, no revocation, and no trail. Here is the self-hosted git forge I put in front of prod instead, the traps that bit during the migration, and the last layer I honestly haven't built yet.

#ai#opsec#self-hosting#git#agent-security
IT & SystemsNov 9, 20266 min

One quant, two machines: Unsloth Dynamic v3.1 across five GPU generations

The same Qwen3.8-Flash-Next quant landed on my Strix Halo and my 1080 Ti on the same day. Same weights, same release notes — completely different optimization problems. Here is the sizing math, the receipts, and the three ops traps I fell into while upgrading both.

#unsloth#quantization#strix-halo#1080-ti#gufo#strata#qwen
IT & SystemsNov 7, 20267 min

The third death notice was a compiler: tuning an inference engine on a 2017 GPU

CUDA 13 stopped compiling for Pascal — the third formal death notice for my 1080 Ti fallback. Before accepting it I ran a proper A/B campaign on the engine: one free 20 % decode win, one flag I kept on bitwise-parity evidence, three I rejected as noise, and two measurements I distrust even now.

#gpu#nvidia#pascal#cuda#local-llm#benchmarking#strata
ElectronicsNov 1, 20266 min

The EVGA that refuses to die: a 2017 GPU as a production fallback

NVIDIA sent Pascal to the legacy driver branch. EVGA itself left the GPU business. The card got two death notices and answered with receipts — including the time I caught my own first draft's numbers being wrong and re-measured everything.

#gpu#nvidia#pascal#local-llm#reliability#benchmarking