A production-ready MLOps platform for sales forecasting that demonstrates modern machine learning engineering practices. Built on Astronomer (Apache Airflow), this project implements an end-to-end ML pipeline with ensemble modeling, comprehensive visualization, and real-time inference capabilities via Streamlit.

• Automated ML Pipeline: End-to-end orchestration with Astronomer/Airflow
• Ensemble Modeling: Combines XGBoost, LightGBM, and Prophet for robust predictions
• Advanced Visualizations: Comprehensive model performance analysis and comparison
• Real-time Inference: Streamlit-based web UI for interactive predictions
• Experiment Tracking: MLflow integration for model versioning and metrics
• Distributed Storage: MinIO S3-compatible object storage for artifacts
• Containerized Deployment: Docker-based architecture for consistency

• Docker Desktop installed and running
• Astronomer CLI (brew install astro on macOS, other OS, you can follow the instructions here)
• 8GB+ RAM available for Docker
• Ports 8080, 8501, 5001, 9000, 9001 available

# Clone the repository
git clone https://github.com/airscholar/astro-salesforecast.git
cd Astro-SalesForecast

# Start Astronomer Airflow services
astro dev start

This will start:

• Airflow UI: http://localhost:8080 (admin/admin)
• Streamlit UI: http://localhost:8501
• MLflow UI: http://localhost:5001
• MinIO Console: http://localhost:9001 (minioadmin/minioadmin)

1. Open Airflow UI at http://localhost:8080
2. Enable the sales_forecast_training DAG
3. Trigger the DAG manually or wait for scheduled run
4. Monitor progress in the Airflow UI

1. Open Streamlit at http://localhost:8501
2. Click "Load/Reload Models" in the sidebar
3. Choose input method (upload CSV, manual entry, or sample data)
4. Configure forecast parameters
5. Generate predictions and export results

• Synthetic data generation with realistic patterns
• Time-based train/validation/test splitting
• Comprehensive data validation and quality checks
• Advanced feature engineering (lags, rolling stats, seasonality)

• XGBoost: Gradient boosting for non-linear patterns
• LightGBM: Fast training with categorical support
• Ensemble: Optimized weighted average of all models
• Hyperparameter tuning with Optuna

• Model performance comparison charts
• Time series predictions with confidence intervals
• Residual analysis and diagnostics
• Feature importance rankings
• Interactive plots with Plotly

• Automated experiment tracking with MLflow
• Model versioning and registry
• Artifact storage in MinIO
• Production model promotion workflow

• Multiple Input Methods: CSV upload, manual entry, sample data
• Model Selection: Individual models or ensemble
• Interactive Visualizations: Real-time prediction plots
• Confidence Intervals: 95% prediction bounds
• Export Capabilities: Download predictions as CSV

# Simplified prediction flow
Input Data → Feature Engineering → Model Prediction → Visualization → Export

• Training Time: ~2-5 minutes for full pipeline
• Prediction Latency: <100ms per forecast
• Model Accuracy: MAPE < 5% on test data
• Ensemble Performance: 15-20% improvement over individual models

1. Services not starting: Check Docker memory allocation (8GB minimum)
2. Models not loading: Ensure training DAG has completed successfully
3. Port conflicts: Stop conflicting services or modify ports in docker-compose
4. MLflow connection: Verify MLflow service is running and accessible

# Check Airflow logs
astro dev logs

# Check specific service logs
docker-compose -f docker-compose.override.yml logs mlflow
docker-compose -f docker-compose.override.yml logs streamlit

• Detailed Architecture
• Astronomer Docs

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests if applicable
5. Submit a pull request

This project is licensed under the MIT License - see the LICENSE file for details.