Azure ML lifecycle
Individual MLOps coursework, 2025
- Role
- Individual implementation
- Stack
- Azure ML, TensorFlow, Python, FastAPI, Docker, GitHub Actions
- Model
- 15-class sports-ball image classifier
A machine-learning workflow that can be rerun from start to finish. Data preparation and training run on Azure ML, and GitHub Actions downloads the registered model, serves it from a container, and checks it automatically.
Lifecycle
Everything around the model
In this course the model was only one piece of the system. Azure resources, data assets, environments, pipeline components, model versions, inference services, and deployment checks were all defined explicitly.
Stage 1Azure data assetsFifteen classes resized to 64×64 in parallel jobs
Stage 2Data split80/20 train-test split within each class
Stage 3Cloud trainingA Keras CNN on an Azure ML compute cluster
↓ metrics + artifacts ↓
Stage 4Model registryVersioned model with its metadata
ServingFastAPI inferencePredictions, a health endpoint, and prediction history in PostgreSQL
DeliveryCI/CD workflowBuild, deploy, health check, and notification
Implementation
What I built
- Automated the Azure setup: resource group, workspace, compute cluster, environments, components, and the data upload.
- Built a four-stage Azure ML pipeline: fifteen independent preprocessing jobs, splitting, CNN training with augmentation and early stopping, and model registration.
- Registered the trained model in Azure ML, so deployment starts from a registered model instead of a local file.
- Built a FastAPI inference service and a Streamlit dashboard with Grad-CAM, backed by a history of predictions in PostgreSQL.
- Containerized the system with Docker Compose. I also wrote Kubernetes manifests for it, but the workflow deploys with Docker Compose.
- Used GitHub Actions to automate Azure setup, training, downloading the model, deployment on a self-hosted runner, health checks, a test prediction, deleting the Azure resources afterwards, and status messages to Discord.
Engineering
From raw data to a running service
- VersioningData assets, environments, and models have versions in Azure ML, and every workflow run tags its container images with the run number.
- Cloud executionPreprocessing and training run on the Azure ML resources I created, not only on a local machine.
- Deployment checksThe workflow checks the service's health, runs a test prediction, and posts the pipeline status to Discord.
- ExplainabilityThe Streamlit dashboard uses Grad-CAM to show which parts of an image influenced a prediction.
- Kubernetes manifestsThe manifests include probes, resource limits, persistent storage, secrets, and two API replicas. I didn't deploy them to a cluster.
- ScopeTraining and model registration ran in the cloud on Azure ML. Serving ran with Docker Compose on a self-hosted machine, not on a live production cluster.