Implement AI to Reduce Mis-Picks Without Replacing Your WMS

Most 10-person warehouse teams rely on their Warehouse Management System (WMS) and handheld barcode scanners to ensure accuracy. Systems like SkuVault, Fishbowl, or NetSuite's WMS module are excellent for tracking inventory levels and bin locations. However, their validation logic ends at the barcode. A scanner confirms a worker picked an item with the correct EAN or UPC, but it cannot confirm they picked the correct *quantity*. It also cannot distinguish between visually identical products with different SKUs, like a blue shirt in size medium versus large, if the wrong barcode was affixed to the bag.

Consider a fulfillment center packing cosmetics orders. A packer needs to put one 'Matte Lipstick - Ruby Red' and two 'Matte Lipstick - Cherry Bomb' into a box. The tubes are identical except for a tiny label on the bottom. The picker scans one 'Ruby Red' and the WMS gives a green light. They then grab what they think are two 'Cherry Bomb' tubes but accidentally take one 'Ruby Red' and one 'Cherry Bomb'. The scanner won't catch this because the initial scan was correct, and the WMS relies on the human to confirm the quantity. The mistake is only discovered a week later when the customer files a return, costing the business in shipping and reputation.

The fundamental issue is that a WMS is a database designed for inventory accounting, not a real-time perception system. Its architecture is built to trust human-and-scanner inputs as the source of truth for physical actions. It has no independent mechanism for verifying that the physical items in a tote match the digital order record. Trying to add this capability directly into a legacy WMS is often impossible without a full-scale, multi-year migration project that a 10-person team cannot afford.

We would begin with a 2-day audit of your WMS API and 12 months of picking error data. The goal is to map WMS order endpoints and identify the most common failure patterns, like quantity mistakes or visually similar product swaps. We would also collect at least 50 images for each of your top 100 SKUs to serve as the initial training set for the computer vision model. This audit produces a clear technical plan and confirms the project's viability.

The core of the system is a Python service built with FastAPI that orchestrates the validation logic. An overhead camera at the packing station sends an image to the service. A computer vision model, likely based on YOLOv5 for its speed and accuracy, detects and counts each item in the image. The FastAPI service then calls your WMS API to fetch the correct order details and compares the vision model's output to the order manifest. The entire process, from image capture to validation result, is designed to complete in under 500ms. All logs and results are stored in a Supabase database for performance tracking.

The delivered system is a simple interface on a monitor at the packing station. After placing items in the box, the packer sees an instant green checkmark for a correct order or a red 'X' detailing the specific error (e.g., 'SKU ABC-123: Expected 2, Found 1'). This provides immediate, unambiguous feedback, allowing correction in seconds. The system acts as an intelligent layer on top of your existing workflow, dramatically improving accuracy without disrupting the WMS you already depend on.