Homelab inference research and infrastructure observations. The method is measure, don't speculate.
Astrophotography is maybe a tenth telescope and nine tenths file management. This is how AI, working as a build partner, helped one hobbyist build and run a small distributed system for it: syncing files safely, planning nights, coordinating four devices, grading frames, and captioning the galleries. The honest version of what AI had to do with it, scars included.
Contextualization was 97.8% of our RAG ingest pipeline. Cutting it from ~25 minutes to under a minute took two separable things: an eval gate honest enough to prove a 3B model was good enough to replace a 31B, and a full-stack hunt — GPU monitor to vLLM metrics to a Postgres commit — that ended at four GPUs sitting idle behind a database transaction.
A citation-heavy drafting eval put self-hosted Gemma 4 31B AWQ on the clean frontier with GPT-5.1. The result is less a leaderboard than a study in claim grounding, citation support, project binding, and model economics.
The expensive failure mode in AI-driven development is not bad code. It is drift: specs, branches, memory, schemas, prompts, and plans aging at different speeds while the model keeps building from stale premises.
What two RTX 3090s, an 8-cell cold-start sweep, and a power-cap experiment taught me about where the seconds and watts actually go. Three assumptions the data didn't support, one 12× ratio that surprised me, and a 250 W power cap that gives back 36% of GPU power for an 11% throughput cost.
Standing up vLLM nightly and llama.cpp on the same 3090 with the same Qwen3.6-27B model — and discovering that two inference engines on identical hardware give a 4× difference in usable context. Hybrid Mamba-attention architecture accounting, quantization comparison, and the prompt-cache mechanics behind the gap.