Artificial Intelligence
Logistics & supply chain
Demand forecasting for a multi-region distributor
16-week engagement · 4 engineers & 1 ML lead
- Challenge
- Planning ran on spreadsheets across three regions with 40,000 SKUs. Warehouses were overstocked in some markets and short in others, and every regional planner forecast in their own way — so leadership had no reliable number to buy against.
- What we built
- A machine-learning forecasting service trained on four years of order history, weather and promotional data, with per-region hierarchy so regional planners see their own numbers. Every recommendation lands in a planning dashboard with human approval, and the model retrains weekly as new orders arrive — accuracy is tracked per category and surfaced to the business, not hidden in a notebook.
PythonTime-series MLCloud data platformPlanning dashboard
32%
lower excess stock
18%
fewer stockouts
6 wks
to first production model
In the end: Within two quarters the buying team shifted from weekly firefighting to exception-based planning — the forecast became the default, and planners spent their time on the outliers the model flagged.
