Ai Solutions
Ai Solutions
A manufacturer with 17 plants across 8 countries faced $42M in annual waste from demand-supply mismatches. Spreadsheet-based forecasting couldn't handle 14,000 SKUs with seasonal patterns.
A multinational manufacturer with 17 plants across 8 countries was bleeding $42M per year to demand-supply mismatches. The forecasting team ran 14,000 SKUs through a patchwork of spreadsheets and regional tools, none of which could reason about cross-plant dependencies or seasonal demand shifts. Stockouts triggered rush production; over-forecasts sat in warehouses for months.
We replaced the spreadsheet layer with a hierarchical forecasting system that reasoned across plants, SKUs, and time horizons simultaneously. The critical design decision was treating the supply chain as a single optimization problem rather than 17 independent ones. We started with the top 200 SKUs (which drove 60% of waste) and expanded once the model was trusted.
Transformer-based hierarchical time-series model with plant and SKU embeddings learned from 4 years of historical demand. Probabilistic outputs feed an inventory optimization layer that solves a multi-echelon replenishment problem. Outputs are pushed to the ERP every 4 hours via a streaming pipeline; planners see ranked recommendations with confidence intervals and override UI for exceptions.
22 weeks end-to-end. Weeks 1-6: data harmonization across 17 plants (the hardest phase). Weeks 7-12: model development and backtesting. Weeks 13-18: ERP integration and planner UI. Weeks 19-22: phased rollout plant by plant, with parallel running.
After 12 months in production: $18M reduction in waste, 71% fewer stockout incidents, 34% lower inventory holding costs. Forecast accuracy improved 41% over the previous baseline. The planning team shifted from producing numbers to managing exceptions.
Want to discuss a similar challenge?
Start a Conversation