Field notes from running production AI on my own hardware, and the privacy engineering behind it. Long-form lives here; day-to-day benchmarks, recipes and community traffic live in the club-3090 Discussions.
Every business is told to adopt AI. Almost nobody asks the first question: where does the data go? What cloud models actually do with your prompts, why agents raise the stakes, what "private" means on a spectrum from VPC to your own GPU box — and an honest decision test for when it's worth it.
~6 min read · no slideware
What private AI actually costs (a TCO that survives a CFO) · agents and data boundaries · the post-training arc — RAG → fine-tuning → RL, worked in public on the reference rig · picking GPUs for local inference · fine-tuning on your own hardware, end to end.
published when measured — not before