Predicting Supplier Delivery Quality with Machine Learning

Supplier disruptions are often visible in operational data before they become obvious to planners, but turning those signals into a useful decision tool requires more than training a model.

For my master’s thesis, I translated stakeholder needs from cross-functional workshops into a supplier-risk use case, analyzed data from an industrial data lake and developed an end-to-end Python and machine-learning solution. To make the results usable, a web dashboard was developed to make the results accessible for management and supplier-management workflows.

The ML pipeline uses pandas and scikit-learn for feature engineering, XGBoost for the risk model, and Optuna for hyperparameter tuning. A FastAPI backend serves predictions from a PostgreSQL database, packaged in Docker and deployed on AWS, with a React and Vite dashboard giving supplier managers a day-to-day view of the risk scores.

The solution identified approximately 25% of subsequent supplier risks up to five days in advance, giving supplier managers an earlier window to act before a disruption reached the shop floor.