I built an open-source AI gateway for Kubernetes: budgets, kill switch and evals in CI A platform engineer built SanVi, an open-source AI gateway for Kubernetes that routes app traffic through a single OpenAI-compatible LiteLLM endpoint with per-team virtual keys, budgets, rate limits and a kill switch. The gateway logs every call's team, model, tokens and cost to Postgres, keeps the provider key in Vault synced via External Secrets, and enforces policy through pull-request-based team configs checked in CI and applied by Argo CD. A Right-Size Advisor runs the same eval suite per model — on the author's workload Haiku matched Sonnet at 95% lower cost and 40% faster — and fails CI on any change dropping below 80%. Every team wants AI. Then three questions arrive: What does it cost? Where are the keys? Can we stop it? I'm a platform engineer, and I built SanVi to answer all three on Kubernetes. Apps call one OpenAI-compatible endpoint LiteLLM engine with their own virtual key. Each key has its own models, budget and rate limit. No key → 401. Over the rate limit → 429. Over budget → blocked. Every call is logged with team, model, tokens and cost in Postgres. The provider key lives only in Vault and syncs in through External Secrets. I rotated it: about 70 seconds, no redeploy. Leaked key, runaway agent loop, spend spike: kill switch.sh block