This site exists because my notes deserve better than a folder of .txt
files named final_v2_REALLY_final.md.
The stack behind the curtain
Everything you're reading runs off hardware I own and administer:
- Strix Halo mini-PC (128GB unified memory) serving a quantized Qwen model at 256k context — it powers my personal AI assistant, 24/7.
- Small virtualization cluster for everything else.
- Zero inbound ports. The whole AI stack is channel-only: outbound SIP and email, no listeners facing the internet.
home rack ── virtualization cluster ──┬── gufo (LLM inference, Strix Halo)
├── assistant gateway (channel-only)
└── this site, eventually
What to expect
Field notes, not tutorials. If you want "Top 10 Linux Commands," there is an LLM farm producing those by the ton somewhere. What you'll find here:
- Real builds with real constraints (budget, noise, power, WAFs).
- Post-mortems of things I broke, with the actual terminal output.
- Occasional detours into electronics and cars, because a soldering iron is just a pen for electrons.
First real post incoming: how the inference box actually performs under daily load, numbers included.