Amir Saleem

End-to-End MLOps Pipeline for Credit Default Prediction

PYTHON • PANDAS • SCIKIT-LEARN • DECISION TREES • GITHUB ACTIONS • SQL

Built an end-to-end MLOps pipeline to predict customer credit defaults using a Decision Tree model, with a strong focus on automation, monitoring, and production reliability.

The system ingests data directly from a database, performs data cleaning and feature engineering, and trains a machine learning model evaluated using F1-score to balance precision and recall.

Automated retraining:
• Periodic retraining (weekly)
• Drift-triggered retraining using Population Stability Index (PSI)

Integrated CI/CD using GitHub Actions to automate model retraining, evaluation against the production model, versioned model storage (.pkl), and deployment to the API server.

Safe deployment: model versioning and rollback; blue-green deployment strategy. Monitored production performance by tracking input data, predictions, and anomalies—identifying and resolving issues such as reduced data ingestion from UI changes and performance degradation from data drift.

The system is designed for scalability, reliability, and continuous model improvement with minimal manual intervention.