Deployment · Hardening · Operations
Self-Hosted AI API Gateway
Deployment & Operations
I deploy, harden and operate self-hosted AI API gateways for teams that want one controlled entry point for their LLM traffic — with their own keys, their own logs, and their own infrastructure.
# every deployment ends with proof, not promises
$ deployctl preflight --remote gateway-host
ok ports 80/443 free · docker 27.x · 4GB free
$ deployctl apply --plan plan.json
ok container healthy · bound 127.0.0.1 only
$ deployctl verify --target gateway-host
ok TLS A · HSTS · headers 6/6 · logs redacted
# handover: runbook + backups + your keys, not mine
The problem
Keys everywhere
Provider API keys pasted into notebooks, CI variables and desktop apps. Nobody knows which key is used where, or what it would cost to rotate one.
No spend visibility
The invoice arrives at month end. Which team, which feature, which runaway script spent the budget? The provider dashboard cannot tell you.
Provider lock-in at the code level
Every service hardcodes one vendor SDK. Switching models — or surviving an outage — means a code change and a deploy.
What a gateway fixes
A self-hosted gateway (open-source projects such as LiteLLM, one-api or Sub2API-class deployments) gives you one OpenAI-compatible endpoint in front of every upstream you use. Your services get virtual keys you can create, cap and revoke; upstream keys stay in one vault. You get per-key usage, rate limits, model routing and failover — on your own server.
What I deliver
Deployment
Docker Compose or systemd deployment, reverse proxy with HTTPS and HSTS, non-root containers, a documented directory layout, and an upgrade path that has been tested, not guessed.
Hardening
Key vaulting with restricted permissions, admin plane separated from the data plane, security headers, fail2ban, log redaction so secrets never land in plaintext logs.
Operations
Health checks and alerting, usage dashboards, backup and restore drills with a written runbook, quota policies per team and per key.
Handover
You own everything at the end: infrastructure, documentation, and a recorded walkthrough. No dependency on me to keep it running.