01
Private AI platforms often wire identity, policy, cost, and SLOs as scattered service configs. That makes ownership, audit, and platform-level governance hard to operate.
Case 26 · Flagship
AI Infra Control Plane
An AI infra control plane connects policy, identity, audit, intent, MCP, FinOps, and SLO signals into one operating model for private AI on Kubernetes.
01
How it works
From signal to a verified proposal.
AI Infrastructure Control Plane by Andrey Lesnikov — Kubernetes-native governance for private AI: Detect → Decide → Approve → GitOps → Verify.
-
01
Evaluate the PolicyBundle
Versioned policy packs score the request against identity, quota, MCP boundaries, FinOps, and SLO inputs — producing a durable decision_id.
-
02
Record a durable decision
Decisions and approvals persist (Postgres HA or SQLite single-node) with tenant isolation and JWT-only tenancy where required.
-
03
Approve and propose GitOps
Capability-bound approvals unlock remediation proposals — including optional GitHub draft PRs — instead of silent config drift.
-
04
Verify the runtime contract
Signed OCI Helm/images, Trivy scans, and runtime verification close the loop so governance is proven, not assumed.
02
Evidence
From problem to proof.
How the system is framed, operated, and proven — not a slide deck summary.
02
The control plane unifies OPA policy, OIDC identity, audit trails, intent routing, MCP tool boundaries, FinOps signals, Redis state, and Prometheus SLO telemetry under Kubernetes ownership.
03
Platform owns PolicyBundle generations, durable decisions, capability-bound approvals, and GitOps proposals. Runtime planes consume signed contracts instead of inventing governance per service.
04
A shared control plane concentrates platform responsibility. The payoff is consistent contracts across AI services instead of every team reinventing governance.
05
Useful signals include decision IDs, approval records, PolicyBundle generations, GitOps proposal status, runtime verification results, and audit trails around MCP and intent.
06
Quota backend fail-closed (PR #125): default QUOTA_ON_UNAVAILABLE is approval_required; unknown values also fail closed. allow is explicit opt-in only and is marked with reason quota_backend_unavailable_allowed plus a warning metric. Compose e2e adds a Redis-down step asserting a non-allow decision; kind e2e waits on real Argo CD Synced status (docs/EVIDENCE.md).
Constraints
OPA policy, OIDC identity, MCP tool boundaries, FinOps signals, Prometheus SLOs, Redis state, and Kubernetes ownership.
03
Related
Nearby systems.
AI Runtime Platform
AI workloads often start as API calls but become production systems that need identity, policy, scaling, telemetry, and clear ownership
Every model call: policy + trace + budget
Cloud-Native AI Gateway
AI usage needs routing, policy, budget awareness, and provider resilience
One boundary: rate, budget, fallback, OTel
RAG Knowledge Platform
Engineering knowledge is spread across repositories, runbooks, tickets, architecture notes, and project history
Cited answers from project evidence
Available for meaningful infrastructure conversations