Engineering notes from building and running a home lab — systems work written up for anyone who finds the problem interesting. Each one is a real thing I built, the approach that didn't work, and the approach that did.
-
Running two LLM brains on two mismatched GPUs — dedicating a model per GPU (vs tensor-splitting one across both), MoE for big-model quality at small-model speed, and keeping a card free for hardware video encode.
-
Killing notification fatigue with a triage layer — per-host observers writing to a filesystem inbox, one coordinator using an LLM to triage now/morning/never, and a hard privacy wall between two users on shared hardware.
-
An automated disc → media-library pipeline — hands-off DVD ripping + hardware transcoding, and the metadata-matching war stories (volume labels, nfo/artwork precedence, audio fingerprinting).