standalone AI on one server

A single small server runs the sites, the API, the workers, the mail and the database. Background work is a table in the database with row-level locks and a few systemd timers; there is no message broker, no container runtime, no dashboard until there is something to watch. Releases are immutable directories with an atomic symlink switch and a health check that rolls back on its own; only the last three are kept, so the disk cannot fill with builds.

Less than people think. Language models are called over the network with a strict timeout; the decision model runs locally in a millisecond; speech runs as a separate process on the same box. The expensive part of an AI product is never the inference, it is the failure states, and those are code.

One server, one small backup bucket, and the per-call prices of the vendors that are genuinely irreplaceable: the phone network, card payments, speech. Everything else is ours.