Private AI deployment architecture for one client, layer by layer: retention, communication, access, model inference, storage and where it runs.