The problem
A fruit exporter's inventory is dying from the moment it arrives. Our client handles on the order of thirty million kilograms of apples a year from 150-plus growers, and every lot in cold storage is a countdown: process it in time and it becomes export product; wait too long and it becomes juice, or loss. Today that triage runs on experience and rules of thumb — and rules of thumb, at thirty million kilos, are expensive.
The solution
Our design replaces the thumb with a model. Machine-learning models — gradient-boosted, trained on the company's own history in its ERP — predict remaining shelf life per lot from variety, grower, harvest data, and storage conditions, and score supplier quality while they're at it.
A prioritisation engine turns predictions into the daily processing queue: which lots first, which can wait, which need quality inspection now. Alerts flag lots drifting toward the edge, and compliance checks match fruit against destination-market requirements before it is committed.
The design is deliberately modest to run — microservices beside the existing ERP, sub-second predictions, no rip-and-replace — and costed in the low tens of thousands of dollars, pocket change against a single season's spoilage at this scale.
