RetailSense AI
A demand-forecasting platform for a 54-store grocery chain, deployed twice over: Docker Compose and Kubernetes.
RetailSense AI is a production-grade MLOps platform forecasting daily demand across 54 stores and 33 product families for Corporacion Favorita, an Ecuadorian grocery chain. Twenty-one tracked MLflow runs compare LightGBM, XGBoost, and an ARIMA baseline; LightGBM wins a paired t-test A/B comparison (p=0.0036) at RMSE 336, MAE 230, MAPE 8.78%, a 55% RMSE improvement over the ARIMA baseline, and is registered as the production model. A FastAPI backend exposes forecast, segmentation (K-Means/GMM), SHAP explainability, A/B-test, and drift-detection endpoints, served through a React and Recharts dashboard behind Nginx. The full stack, FastAPI, React, MLflow, PostgreSQL, MinIO, was deployed twice over on the same AWS EC2 instance: once via Docker Compose and once via Kubernetes (Minikube). That EC2 instance has since been stopped, the project is presented here through its GitHub repo and screenshots.
What I built
- Built a full feature-engineering pipeline (lag features, rolling statistics, date decomposition, holiday and oil-price encoding, store clustering) over 4+ years of daily sales across 54 stores and 33 product families.
- Tracked 21 MLflow runs across LightGBM, XGBoost, ARIMA, K-Means, and GMM, selecting the production model through a paired t-test on MAE (p=0.0036) rather than a single-run comparison.
- Registered LightGBM (RMSE 336.33, MAE 229.67, MAPE 8.78%) as the production model, a 55% RMSE improvement over the ARIMA baseline (RMSE 740.71).
- Exposed forecast, segmentation, SHAP explainability, A/B-test, and KS-test drift-detection endpoints through a FastAPI backend, with a React and Recharts dashboard behind Nginx.
- Deployed the same 5-service stack (frontend, backend, MLflow, MinIO, PostgreSQL) twice: via Docker Compose and via Kubernetes (Minikube) on the same AWS EC2 instance, with NodePort services exposing each Kubernetes deployment.
- Built a KS-test based drift monitor comparing incoming request feature distributions against training data.
Tech stack
Key results
336.33
RMSE (LightGBM)
8.78%
MAPE
55%
RMSE improvement vs ARIMA
21
MLflow tracked runs


![Forecast prediction: 2,486 units, 95% CI [1,827, 3,146], LightGBM](/_next/image?url=%2Fscreenshots%2Fretailsense-ai%2Fforecast-result.jpg&w=3840&q=75)


