Some work should not leave your control, and the usual response is to accept a weaker product in exchange for that. This is what I built to find out how much of a compromise it actually is.
A single machine runs open models locally, with no external API and no data leaving the device. On top of it sits a set of always-on automations: inbox triage, a scheduled daily briefing, and retrieval-augmented answering over a private knowledge base.
How it is built
The design separates a planner from the workers. A lightweight brain decides what should run and when. The local inference server and a handful of small, single-purpose applications do the work, scheduled by the operating system and stored locally.
- Local inference for open weights, served on-device.
- A private retrieval index over my own documents, so answers come from my material rather than from the general internet.
- Scheduled automations that run unattended and report in.
Why it is here
It is the evidence behind a position I take with clients: the private end of the deployment curve is not a penalty box. With current open weights and a disciplined architecture, an organization gets real daily leverage from AI while keeping sensitive data on infrastructure it owns outright. This appliance is also where the private retrieval corpus behind the rest of the estate lives, so the claim is load-bearing rather than illustrative.