Adriel's Lab > Homelab

Single role per node, keys not passwords, and nothing exposed to the open internet. Weak hardware stays stable doing one thing — that’s the whole philosophy.

• The Mesh Seven nodes on a private WireGuard-based network — every service reachable from anywhere I am, none of it reachable from anywhere else. 7 nodes · key auth only · 0 open ports
• Monitoring Prometheus, Grafana, and disk-health tracking on a dedicated node — a repurposed 2010 gaming laptop that now watches everything else work. Prometheus + Grafana · SMART tracking · one pane of glass
• Storage & Backup NAS-primary storage with a nightly tiered backup chain across machines, and a salvaged OptiPlex keeping a second copy of the family archives. A file has to try very hard to die here. NAS primary · nightly chain · tiered copies
• Bitcoin Node A pruned node on a fanless-quiet box — verifying the chain for myself, because that’s the point. pruned · self-verified · local-only
• GPU Node An RTX 3090 that earns its keep: Whisper transcription, embeddings, image ML, and every local LLM in the lab. Operated under one hard rule — one big tenant at a time. Also where I found out that 24GB of VRAM is rookie numbers once a full agentic workflow is involved. 24GB VRAM · ~17× realtime Whisper · one tenant at a time
• Photo ML Offload Immich machine-learning jobs routed from the NAS to the GPU node, so face and object search stay local and fast. Immich · GPU offload · local-only
• Triple-Boot Laptop Windows, CachyOS, and Gentoo on one machine — the btrfs shrink math and the physical-drive-removal trick that stops Windows from hijacking the wrong bootloader. 3 OSes · 2 drives · 0 bootloader collisions

The machines

What the roles above actually run on. Every figure below was read off the machine itself, not from a receipt — core counts are threads, memory is installed capacity. Most of this hardware is old, cheap, or salvaged, which is the point.

• Main workstation The GPU node. Everything heavy runs here — Whisper, embeddings, every local model — and it is the only one that was bought new for the job. 24-thread desktop · 24GB VRAM · bought new
• The NAS A Zettlab D6 Ultra — the odd one out here, since it wears out its own boot SSD running the AI software it ships with. Mine arrived nearly perfect, which makes me one of the lucky Kickstarter backers. The source of truth, and deliberately not used as AI compute. 18 threads · iGPU only · storage-first
• Dell OptiPlex Saved from a roach infestation left behind by old tenants, now doing archive duty and orchestration. A 2011 CPU is entirely sufficient for scheduling things that happen at 4am. Core i5-2400 · 4 threads · 16GB · no GPU
• A small fanless mini-PC The Bitcoin node. Quiet, pruned, and doing exactly one job — 25GB used of a 917GB disk. 4 threads · 16GB · 3% disk used
• Alienware M15x A 2010 gaming laptop with a dead battery, now the monitoring node. It watches ten other exporters and has no business still being useful. Core i7 Q 740 · 8 threads · 8GB · 1.73GHz
• Raspberry Pi 4 Model B The gateway — exit node, subnet router, and recursive DNS. The smallest machine here and the one everything else reaches the outside world through. Pi 4B rev 1.5 · 4 cores · 8GB · gateway + DNS
• Triple-boot laptop The portable one, and the only machine running three operating systems off three drives. 6GB of VRAM makes it the mobile tier — small models locally, and everything bigger reached back over the mesh. 16 threads · 32GB · 6GB VRAM · mobile tier
• Two borrowed gaming rigs The only real burst capacity here, and the only machines on this page I don’t own — they belong to family and are gaming PCs first. They get used for parallel transcription when they happen to be idle, and never get planned around. 2 machines · 12GB-class NVIDIA each · burst only Specs kept vague on purpose — these machines aren’t mine to itemise.