Course outline for Platform Engineering for AI Teams

Module 1: Define an AI platform contract

• Internal developer platforms: Define AI developer users, model/data owners, supported providers, and operating budgets; make model access a maintained platform contract rather than a shared credential.
• Infrastructure as code: Version environment declarations and validate them against a local contract before producing a change preview for approval.
• Terraform: Inspect a supplied local-target plan and distinguish its proposed changes from infrastructure state and approved provisioning actions.

Practical outcome: Define model/data owners, approved provider options, and budget settings; inspect a Terraform plan for the supplied local mock AI environment.

Module 2: Make AI delivery repeatable

• GitOps: Trace desired-state reconciliation from supplied records, separating convergence of infrastructure from pipeline checks and developer request completion.
• CI/CD: Gate mock AI delivery on contract and supplied evaluation checks with pinned prompt, dataset, and application versions; passing checks still requires approval.
• Developer environments: Pin runtime, model configuration, and dataset fixtures; reject missing versions and unsupported provider/data combinations before mock environment provisioning.

Practical outcome: Trace desired-state reconciliation and run supplied contract/evaluation-fixture gates with pinned prompt, dataset, and application versions before environment approval.

Module 3: Support AI self-service and exceptions

• Golden paths: Offer a supported AI application template with model access, data policy, telemetry, and test hooks; keep an owned exception route for unusual workloads.
• Self-service infrastructure: Provision from an approved preview with request-key deduplication, requiring renewed approval when preview parameters change.
• Exception routes and platform feedback: Route nonstandard requests to named owners and measure waiting, completion failures, and support touches before proposing a reviewed improvement.

Practical outcome: Provision an approved mock environment, reject unsupported provider/data combinations, deduplicate requests, and route model or capacity exceptions to a named owner.

Practical exercise

• Produce a versioned AI environment template, validation checks, preview record, and ownership/exception contract using supplied infrastructure and evaluation scaffolds.
• Accept a valid request; reject missing dataset versions, forbidden provider/data combinations, and a failed evaluation-fixture gate before provisioning.
• Approve a preview and run the mock provisioner twice; verify one environment per request key and renewed approval when provider or budget parameters change.
• Record an unsupported model/capacity request and measure developer waiting, failures, and support touches; propose one reviewed improvement with separately controlled credential delivery.