E-Commerce Customer Risk ML Pipeline and API
Python · DuckDB · XGBoost · Optuna · SHAP · MLflow · Azure Machine Learning · FastAPI
Overview
This project demonstrates an end-to-end analytics and ML workflow using the Brazilian Olist e-commerce dataset: SQL-based feature preparation, statistical analysis, model tuning and explanation, experiment tracking, and an API that returns a risk score with an optional generated retention message.
Pipeline
- Read Olist CSV data and expose it as DuckDB views.
- Derive delivery, order, and product features with SQL.
- Compare review scores for on-time and delayed orders and prepare a classifier dataset.
- Tune an XGBoost model with Optuna, evaluate it using ROC AUC, and generate SHAP explanations.
- Log experiment metrics and artifacts with MLflow, using Azure Machine Learning tracking when configured and local MLflow tracking as a fallback.
- Serve predictions through FastAPI and Pydantic. For high-scoring cases, the API can request a suggested retention message from OpenAI.
Model Target and Limitations
Despite the project’s retention and churn framing, the training target in the code is a low customer review score (1–2 stars). It is a proxy for an adverse customer experience, not a label showing that a customer actually stopped purchasing. The score therefore should not be presented as a validated churn probability.
The API generates message text; it does not send a campaign to customers. The project demonstrates a technical workflow and does not establish business impact or production suitability by itself.
Technologies
Python, DuckDB, SQL, Pandas, NumPy, SciPy, scikit-learn, XGBoost, Optuna, SHAP, MLflow, Azure ML SDK, FastAPI, Pydantic, Uvicorn, OpenAI API, Docker, Render, GitHub Actions.