AI-Powered Developer Platform on ARO

Welcome to the AI-Powered Developer Platform on ARO workshop.

The story for today is a Monday at a company that will not send source code to a public LLM. This morning the platform team publishes Qwen3.8. This afternoon a developer ships an app with it. You wear both hats, then you look at the GPU and token bill.

Who This Workshop Is For

Persona Role in This Workshop

Platform Engineer (morning)

Provisions GPU capacity, installs operators, publishes Qwen3.8 behind a governed OpenAI-compatible API, and hands developers a short AGENTS.md.

Developer (afternoon)

Launches OpenShift Dev Spaces, talks to OpenCode backed by that in-cluster LLM, ships a real app through Tekton and Argo CD, then asks the same model to audit the manifests.

What You Will Build

In-cluster coding agent on ARO

The LLM writes the app. Qwen3.8 is a platform capability.

Act 1 — Platform

  • A fixed-size GPU node pool on ARO HCP — one NVIDIA A100 80 GB, no virtual machines built by hand

  • OpenShift AI Models-as-a-Service serving Qwen3.8-27B (Red Hat AI INT4) behind one HTTPS front door and one sk-oai- key

  • OpenShift Dev Spaces pointed at that gateway

  • AGENTS.md — the only runbook OpenCode reads

  • OpenShift GitOps (Argo CD) and OpenShift Pipelines (Tekton)

Act 2 — Developer

  • A personal Git repository on ephemeral Git (workshop-scoped, no GitHub account)

  • A Fortune Cookie Go app generated live by OpenCode + Qwen3.8

  • Image build and GitOps deploy from git, ephemeral Git, and oc — no Ansible, no local Docker

  • Platform rules (limits, probes, SCC, TLS) audited by the same model

Act 3 — The bill

  • NVIDIA DCGM GPU metrics and vLLM token metrics in the same console

  • MaaS subscription quota — how the platform team caps cost

Learning Outcomes

After completing this workshop, you will understand:

  • How an ARO HCP node pool gives you GPU capacity without building Azure VMs by hand

  • How OpenShift AI (LLMInferenceService + Models-as-a-Service) is a governed OpenAI-compatible API on your cluster

  • How AGENTS.md shifts platform governance from documentation to something the model enforces

  • How Dev Spaces + OpenCode is an AI-augmented inner loop with no laptop install

  • How Tekton and Argo CD form a GitOps path from a single repository

  • How GPU watts and token windows are the real cost of an in-cluster coding agent

Workshop Structure

Act Lab Topic Duration

0 — Get in

—

Deploy the ARO HCP cluster and log in with oc

45 min

1 — Stand up the agent

1.1

Deploy a GPU Node Pool

15 min (5 active)

1.2

Install Platform Operators

5 min active / ~20 min background

1.3

Publish Qwen3.8 as a Service

15 min active / model loads in background

1.4

Hand the Keys (AGENTS.md + Dev Spaces URL)

10 min

2 — First day

2.1

Launch Your Dev Space

10 min

2.2

Prompt to Production (generate, ship, change)

35 min

2.3

Audit with OpenCode

10 min

3 — The bill

3.1

GPU and Token Metrics

15 min

—

—

Conclusion & Q&A

10 min

Fire-and-forget timing: GPU node, operators, and model load are triggered early and run in the background. By the time Sam opens Dev Spaces, Qwen3.8 is Ready.

Prerequisites

You will need:

  • A web browser

  • An empty Azure subscription. Act 0 can run in a local terminal, or from the browser launcher, which starts an Azure CLI terminal in the lab resource group

  • This workshop repository (cloned in Act 0) — not the Go app template

Let’s get started.