This flags orders running late against their promised delivery date, which is the only date 'late' really means. I first tried a model on transit-time percentiles, but p95-of-transit barely correlated with promise-date lateness, about 5.8% precision, so I threw it out and measured against the promise directly instead. It works in two tiers: one for orders already past their promise and not delivered, and one for packages so overdue they're probably lost and need a human. Sometimes the simpler, non-ML answer is just the right one.
The Feed builds, warehouses, and ships fuel for endurance athletes, and it's growing fast. I work on the operations software that lets it scale, the reporting leadership trusts, the forecasts purchasing orders against, and the fulfillment logic that keeps orders on time. I like this kind of software because the result is physical: boxes out the door, hours saved, a shelf that doesn't run empty.
Browse the full collection →