Why my AI stack lives at home

Most of what I’d want an AI assistant to know about me — notes, projects, mail, research — is exactly what I don’t want sitting on someone else’s servers. So the assistant comes to the data instead of the other way around.

The shape of it

Two GPU boxes do the thinking. Dexter carries the main model and every private document; Gadget handles vision and steps in if Dexter goes down. In front of both sits a LiteLLM gateway: every chat window, agent and n8n workflow talks to that one endpoint, which means one place for keys, routing and guardrails.

The rules

• Private data never leaves hardware I own.
• No personal information goes to a cloud model.
• Work and patient data stay out entirely.
• Speed matters more than size — a fast 20B model I’ll actually use beats a slow giant I won’t.

The full drawing is on the AI Stack page.