Module 4 — Conclusion & What’s Next

Congratulations! 🥠

You have completed the AI-Powered Developer Platform on ROSA workshop. In under three hours, you deployed an end-to-end AI developer platform and shipped a real application through it.

What You Built

As a Platform Engineer

Component Technology

Component	Technology
GPU infrastructure	ROSA managed machine pool (fixed 1 node), NVIDIA L40S
LLM as a Service	OpenShift AI (RHOAI) + KServe + vLLM + Qwen3.6-35B-A3B-AWQ
AI-native governance	`AGENTS.md` — machine-readable platform conventions enforced by OpenCode
Developer enablement	Go app template + OpenShift Dev Spaces pre-configured with OpenCode
GitOps delivery backbone	OpenShift GitOps (Argo CD) + OpenShift Pipelines (Tekton)
LLM observability	NVIDIA DCGM dashboard + OpenShift AI model metrics (Prometheus)

GPU infrastructure

ROSA managed machine pool (fixed 1 node), NVIDIA L40S

LLM as a Service

OpenShift AI (RHOAI) + KServe + vLLM + Qwen3.6-35B-A3B-AWQ

AI-native governance

AGENTS.md — machine-readable platform conventions enforced by OpenCode

Developer enablement

Go app template + OpenShift Dev Spaces pre-configured with OpenCode

GitOps delivery backbone

OpenShift GitOps (Argo CD) + OpenShift Pipelines (Tekton)

LLM observability

NVIDIA DCGM dashboard + OpenShift AI model metrics (Prometheus)

As a Developer

Milestone Details

Milestone	Details
Personal Git repo	Created in seconds via the Git server — no GitHub account, no SSH key setup
Fortune Cookie app	Generated live by OpenCode + Qwen3.6 — `main.go`, `go.mod`, `Dockerfile`, all from a single prompt
Compile gate	`CGO_ENABLED=0 go build -buildvcs=false` verified code before pushing — zero broken builds
Container image	Built by OpenShift Pipelines (Tekton): `git-clone + buildah` — no local Docker required
GitOps deployment	Developer-owned Argo CD instance synced `deploy/base/` → app live at a public HTTPS route
Compliance audit	OpenCode reviewed and fixed resource limits, probes, security context, TLS — all per `AGENTS.md`

Personal Git repo

Created in seconds via the Git server — no GitHub account, no SSH key setup

Fortune Cookie app

Generated live by OpenCode + Qwen3.6 — main.go, go.mod, Dockerfile, all from a single prompt

Compile gate

CGO_ENABLED=0 go build -buildvcs=false verified code before pushing — zero broken builds

Container image

Built by OpenShift Pipelines (Tekton): git-clone + buildah — no local Docker required

GitOps deployment

Developer-owned Argo CD instance synced deploy/base/ → app live at a public HTTPS route

Compliance audit

OpenCode reviewed and fixed resource limits, probes, security context, TLS — all per AGENTS.md

The Key Insights

1. ROSA makes GPU infrastructure a Platform Engineer concern, not a developer concern. A single rosa create machine-pool command provisions NVIDIA hardware on AWS. Developers never interact with EC2 or node selectors.

2. OpenShift AI makes LLM serving a catalog item, not a project. oc apply -k deploying an InferenceService gives every team in the cluster an OpenAI-compatible API — the same interface they already know, running on infrastructure they control.

3. AGENTS.md shifts governance from documentation to code. Platform conventions encoded in a Markdown file are read and enforced by the LLM at generation time. The developer gets a compliant app without opening a checklist.

4. The same LLM can be used as a developer tool AND as an application service. OpenCode used Qwen3.6 to write the Fortune Cookie app. The Fortune Cookie app uses Qwen3.6 to serve fortunes. The platform team deploys one model; it delivers value in both directions.

5. Observability closes the loop. GPU utilization, token throughput, and TTFT are visible in the same console as all other cluster metrics. LLM consumption is measurable and attributable.

What We Did Not Cover (But You Can Explore)

LLM autoscaling with KEDA: The workshop uses a fixed single replica for predictability. In production, proper LLM autoscaling should be driven by pending request queue depth using KEDA with vLLM Prometheus metrics — CPU/memory-based HPA is not meaningful for GPU inference workloads.
Multi-environment GitOps: Add deploy/overlays/stage/ and deploy/overlays/prod/ with Kustomize overlays; use separate Argo CD Application objects per environment
Webhook-triggered pipelines: Add an EventListener + TriggerBinding to Tekton so git push triggers builds automatically
Application autoscaling: Add an HPA to the Fortune Cookie Deployment to scale based on HTTP requests
Model fine-tuning: OpenShift AI Data Science Projects support fine-tuning workflows on proprietary data using distributed training
ROSA HCP (Hosted Control Planes): Isolates the control plane from worker nodes for stronger multi-tenancy and faster cluster provisioning
OpenShift Service Mesh: Add Istio-based traffic management, mTLS, and distributed tracing for the Fortune Cookie → LLM call path

Get Started with ROSA

Talk to Us

If you have questions or want to explore ROSA for your organization, our team is here to help. Use the blue hat icon in the Red Hat Hybrid Cloud Console to start a conversation.

Thank you for attending the AI-Powered Developer Platform on ROSA workshop!

May your deployments be swift, your models be accurate, and your fortunes be wise. 🥠