End-to-End MLOps pipeline

Heart Disease Prediction is an end-to-end MLOps pipeline for clinical heart disease classification, covering the complete production ML lifecycle from experiment tracking and explainability to fairness auditing, drift detection, and observability. The model is trained on 14 clinical attributes using Logistic Regression, with MLflow managing experiment tracking and model versioning. SHAP LinearExplainer provides post-hoc explainability by surfacing feature-level contributions and identifying false negative patterns, while Fairlearn audits demographic fairness across age groups using Demographic Parity and Equalized Odds metrics. Input drift is continuously monitored via the Kolmogorov-Smirnov test, comparing live inference distributions against the training baseline across all features.

The FastAPI inference service is containerized with Docker and deployed on GKE with Horizontal Pod Autoscaling, scaling dynamically under load. Per-sample inference logging captures inputs, predictions, probabilities, and latency metrics for full observability, with average latency benchmarked at sub-millisecond levels. Load testing under 2000 concurrent connections via wrk validated the system's production readiness. Every push to main triggers a GitHub Actions CI/CD pipeline that builds, pushes to GCR, and redeploys to the cluster automatically, ensuring continuous delivery at scale.