Fiftify: computer vision for warehouse receiving

Receiving was the longest operation of the day at Fiftify customers' warehouses: a worker went through the pallet by hand, hunted for each item in the list and typed the quantities in. On a large shipment that runs 15-20 minutes, and it is the single biggest source of inventory error: boxes missed, lookalike SKUs mixed up, quantities eyeballed.

A receiving mistake then propagated across every sales channel for weeks. Shopify advertised stock that wasn't physically there, while what did sit in the warehouse never sold at all.

We built mobile receiving with on-device recognition: the phone camera streams frames to a YOLOv8 model fine-tuned on the customer's own product photos, which detects packages, classifies the SKU and counts units in frame.

The result is reconciled against the expected shipment in Fiftify. Matches are confirmed with a single tap, mismatches are highlighted and require a human decision. The model never accepts a shipment on the worker's behalf; it only removes the manual sorting.

It is exported to ONNX and runs on the device, so receiving doesn't depend on Wi-Fi quality in the far corner of the warehouse. Unrecognised frames go into a labelling queue and feed the next retraining round.