Four fulfillment nodes, two years of operations. Relational and document sources consolidated into one star schema, with a workload forecast and NLP signals from supervisor shift notes.
Units picked per day, 28-day trailing average. Peak season lifts every node.
Each node-day bucketed against that node's own average volume. Scale is fitted, not zero-based.
Held-out final 90 days. Gradient boosting on lag and calendar features.
Lower MAE is better. Both baselines are what a planner would use without a model.
Average seconds per pick task — the slotting opportunity.
Share of notes by cluster label, from TF-IDF + LSA + k-means.
Weekly, one point per node-week.
Freight cost per shipped unit, with delivery performance.
Rules evaluated on every load. Failures are quarantined, never silently dropped.
Recursive forecast fed back into itself, converted to picker-hours at each node's recent units-per-hour rate.