Pricing
Log inGet Started
AI Platform Engineering
Phase 1Weeks 3-411 machine-verified checks · 1 tutorial pack

The AI Control Plane

The model proposes, the platform decides.

Why this phase exists

Every serious AI incident is the same shape: a probabilistic system was given a deterministic privilege. The control plane is the layer that keeps the model advisory and the platform authoritative, and it is the single highest-leverage thing most teams are missing.

What you learn

  • Policy and validation layers between model output and real systems
  • Approve and review gates, and how to design human-in-the-loop without it becoming rubber stamping
  • Audit trails: who asked, what the model proposed, what actually ran
  • Failure containment: blast radius, rollback, kill switches
  • Designing for the day the model is confidently wrong
Outcome: You can put an untrusted model safely in front of systems that matter.

Tools you will actually touch

OPA / policy enginesKubernetes admissionStructured audit logsFeature flags

The assignment pack

Scenario

An internal ops agent can restart services and change config. Legal needs an audit trail, the CTO needs a kill switch, and the on-call engineer needs the agent to never touch the payment service. Build the control plane that makes all three true.

Definition of done
  • Policy layer that denies out-of-scope actions by default, with tests
  • Approval gate on any destructive action, bypassable only by a named human
  • Audit record for every proposal: request, proposal, decision, outcome
  • A kill switch that stops all agent actions within seconds, proven live
How it is graded
Fail

The agent is bounded by prompt instructions only. Nothing enforces the boundary.

Pass

Policy denies by default, destructive actions require approval, and every action is auditable.

Strong pass

Blast radius is explicitly modelled, the kill switch is tested under load, and denied actions degrade gracefully instead of erroring.

The tutorial pack

Put a policy layer in front of an agent

  1. 1Stand up a minimal ops agent with two tools: one safe, one destructive
  2. 2Insert a policy engine between the agent and the tools
  3. 3Write deny-by-default rules and a narrow allow list
  4. 4Add an approval gate and route it to a human channel
  5. 5Wire the audit sink and prove every decision is recorded
The failure you will hit on purpose

Prompt-inject the agent into attempting a destructive action it was told not to take. Watch the prompt-level instruction fail and the policy layer hold. That contrast is the whole lesson.

What you measure at the end
  • Percentage of out-of-scope attempts denied (target: 100 percent)
  • Time from kill-switch flip to last agent action
  • Every action traceable end to end in the audit log

Per-target lab variants

Same outcome, five stacks. You train on the one you will actually run.

AWSIAM boundaries + EventBridge for approvals + CloudTrail as the audit sink
GCPIAM conditions + Cloud Audit Logs + Pub/Sub approval routing
AzureAzure RBAC + Azure Policy + Monitor/Log Analytics audit trail
On-premOPA + Kubernetes admission control + your own append-only audit store
AgnosticOPA sidecar, any conformant K8s, OSS audit sink, no managed lock-in

You must be able to answer

AI use is mandatory on every assignment. The integrity mechanism is defend-your-work, so these are the questions that decide whether the work counts.

  • “Your policy engine is down. What happens to the agent, and is that the right default?”
  • “How do you stop approval gates from becoming rubber stamps?”
  • “What is the blast radius of your worst-case agent action today?”
Previous
Phase 0 · Agentic Foundations
Next
Phase 2 · Production RAG

Want to be walked through this phase?

The packs are delivered with live teaching, reviewed assignments, and a defence.

Train with me