Representative Project
Manufacturing & Supply Chain
Forecasting planners could finally trust
Problem
Demand planning ran on a spreadsheet refreshed once a week. Forecast error showed up as both stockouts and excess inventory, and when a number changed, nobody could say why.
Approach
We built a forecasting service with model selection at the SKU level, backtested against two years of history, and an explanation layer that surfaces the drivers behind every revision — so a planner sees the "why," not just the "what."
Stack
Python, a mix of gradient-boosted and classical time-series models, a modern cloud data warehouse, and a lightweight review interface.
Outcome (illustrative)
The forecast cycle moved from weekly to daily. Planner time shifted from searching for exceptions to reviewing the ones the system had already found.
