Small Thai restaurants often plan ingredients by intuition, and their POS shows past sales but never what to prepare next. MunchBox works with any POS: staff drop the printed receipt into a box that scans it, OCR turns it into data, each dish is broken down into its ingredients, and a Bayesian model forecasts how much of each to prepare for the days ahead. A CMKL University project (URD-102) advised by Sally E. Goldin.
My role
What I built
Built the prediction dashboard and reporting pages in Next.js and React.
Built FastAPI REST endpoints for recipes, ingredients, staff and CSV import/export on MySQL with SQLAlchemy and Alembic.
Built the receipt intake pipeline with PaddleOCR.
Improved and tracked the accuracy of the Bayesian Structural Time Series (BSTS) forecasting model built with PyMC.
Deployed the system with Docker, Nginx and GitHub Actions CI/CD.
Highlights
What we shipped
Prediction dashboard and reporting pages in Next.js
FastAPI endpoints on MySQL with SQLAlchemy and Alembic
Receipt intake pipeline with PaddleOCR
BSTS forecasts in three stock strategies: conservative, balanced and aggressive (50th, 75th and 95th percentile)
