Building a Full MLOps Pipeline | MNIST + FastAPI + MLflow
In this video, I’ll walk you through a complete MLOps pipeline built on the MNIST digit classification problem.
We go beyond just model training and cover the entire lifecycle:
✅ Data Preprocessing – MNIST dataset, PyTorch + Spark integration
✅ Model Training – CNN baseline + CNN+SVM hybrid ensemble
✅ Experiment Tracking & Registry – MLflow for parameters, metrics, and artifacts
✅ Drift Detection – Detect data distribution shifts with Chi-square test
✅ Automated Retraining – Simulating new data arrivals with partitioned Parquet + auto pipeline retrain
✅ FastAPI Deployment – REST API with /predict, /health, and /reload endpoints
✅ Dockerized Serving – Production-ready containerized API
✅ Results & Comparisons – Loss/accuracy curves, confusion matrices, CNN vs CNN+SVM predictions