Azure
AI & ML
advanced
ML Training and Serving — Azure Architecture Template
ML workspace trains on lake data, packages models into a private registry and serves them from AKS behind API Management.
Why this architecture works
- Training and inference are separate planes: experiments never compete with production latency
- Models ship as versioned container images, so serving rollback is an image tag change
- Feature data comes from the governed lake, keeping training and scoring inputs consistent
- API Management wraps the scoring endpoint with auth, quotas and canary-friendly routing
- Workspace credentials sit in Key Vault and inference latency is tracked in Application Insights
What's inside (8 resources)
Datasets
storage-account
Feature Lake
data-lake-gen2
Model Images
container-registry
ML Workspace
machine-learning
Inference Cluster
kubernetes-service
Scoring API
api-management
Secrets
key-vault
APM
application-insights
From template to running infrastructure
- Open this template in the CloudForge visual designer (free account, no credit card).
- Customize resources, names, and connections on the drag-and-drop canvas — or ask Vani, the AI architect, to adapt it.
- Generate production-ready Terraform, ARM, or Bicep in one click.
- Review the plan diff and security scan, then deploy with human approval.