A personal AI stack should be understandable before it is impressive. This note records the boundaries I use when assembling one: a single entry point, explicit internal services, and checks that exercise the real application path.
Start with the failure boundary
The public surface is intentionally small. Caddy terminates HTTPS and routes requests to the internal gateway; databases, caches, and proxy components stay on the private Docker network.
Make health checks useful
A process being alive is not the same as the service being useful. Checks include PostgreSQL readiness, Redis PONG, the local CPA endpoint, and an unauthenticated API probe that should return 401.
Keep the system reversible
Backups include configuration and data needed to restore the gateway. The standby node is kept cold rather than spending resources on artificial load.