Problem
Demand planning leaned on gut feel and a forecast nobody believed. The previous model lived in a notebook, ran when someone remembered to run it, and offered no explanation for any given number — so planners quietly overrode it, and the overrides became the real plan.
Approach
We built a demand model designed to earn its place in production: measured against the planners' own baseline, served behind a fast API, and wired directly into the tool they already use.
- Honest evaluation — we report where the model is weak, not just where it wins.
- Explanations next to every prediction, so a planner can see why.
- Automated weekly retraining with monitoring that flags drift before it bites.
Result
Stockouts fell by nearly a third, inference stays comfortably under 150ms at the p95, and — the part that matters most — the planners trust it enough to plan around it.