A self-hosted support-triage pipeline: a local Ollama model routes tickets and drafts KB-grounded replies for human approval — every prompt change gated by an eval corpus.
Support inboxes are repetitive but not quite scriptable — routing and drafting need judgment. Cloud AI could do it, but piping a support inbox through someone else's API means paying rent on your own mail, forever. A home rig can run the model.
A self-hosted n8n pipeline: inbound support email is triaged by a local Ollama model, which routes each ticket and drafts a reply grounded in a knowledge base. Nothing sends automatically — drafts land in an approval queue for a human. A live dashboard shows tickets, stats, and the pending-drafts count, with a daily digest email.
No prompt change ships on gut feel. A 40-case eval corpus gates the pipeline; AI-primary routing was promoted only when it beat the rules baseline — 88% against 70%. Ops is built for unattended running: exactly-once email processing, a keep-alive task at logon, nightly backups, and a stress-test suite that proves the recovery behaviour.