End-to-end ML pipeline with MLOps for pharmaceutical sales forecasting
An enterprise-grade machine learning system developed for Rossmann Pharmaceuticals to predict store sales six weeks ahead. The solution integrates ZenML for pipeline orchestration, MLflow for experiment tracking and deployment, and a hybrid modeling approach (Random Forest, XGBoost, LSTM). Designed for scalability and reproducibility, it supports inventory optimization and business planning at scale.
Rossmann Pharmaceuticals manages over 3,000 drug stores across 7 European countries. Their finance and operations teams required accurate multi-week sales forecasts to optimize inventory, staffing, and supply chain costs. Existing tools lacked predictive power and reproducibility, prompting the need for a robust, MLOps-enabled forecasting system.
End-to-End Pipeline: Modular ML pipeline with ZenML covering ingestion, preprocessing, feature engineering, training, and deployment
Multi-Model Ensemble: Combines LSTM, Random Forest, and XGBoost for robust forecasting
MLOps Integration: Automated experiment tracking, model versioning, and deployment with MLflow
Advanced Feature Engineering: Captures temporal trends, promotional campaigns, holidays, and store metadata
Real-time API: FastAPI endpoints for serving predictions in production with Docker-based deployment
Business Dashboard: Streamlit application providing visual insights for executives and store managers
Pipeline orchestration with ZenML ensuring reproducibility and modularity
Data preprocessing for missing values, outliers, and categorical encoding
Feature engineering: lag variables, rolling statistics, promotional indicators, holiday/event features
Hyperparameter tuning with GridSearchCV and model selection based on RMSE, MAE, and MAPE
Ensemble forecasting by blending statistical, tree-based, and deep learning models
Deployment with Docker, MLflow Model Registry, and continuous delivery pipelines
Delivered 92% accuracy in 6-week sales forecasting. Rossmann leveraged the system to optimize inventory levels, reduce operational costs, and improve supply chain resilience across 3,000+ stores.
Explore the complete implementation, documentation, and code on GitHub.