All projects
March 2025
RetailForecastAI — MLOps Demand Forecasting
MLOpsTime SeriesData EngineeringGCPCI/CD
- Production forecasting system that predicts demand across 1,115 store-product combinations at once, then explains each forecast in plain English instead of a spreadsheet of numbers.
- End-to-end MLOps: Apache Beam ETL → PySpark feature engineering (rolling lags, calendar signals, seasonal decomposition) cuts feature-prep time 55%.
- BigQuery ML ARIMA_PLUS trains 1,115 per-store models with 80% confidence intervals, backtested against seasonal-naive baselines at MAE 1,898.84 and MAPE 16.92%.
- RAG narrative layer (LangChain + FAISS + LLaMA/Groq/OpenAI) grounds executive summaries in retrieved facts; Next.js dashboard pairs D3.js charts with a Three.js/WebGL interactive 3D scene and Tableau backtest reporting.
PythonApache BeamPySparkBigQuery MLFastAPILangChainFAISSThree.jsD3.jsTableauDockerGCP