Build an AI System for Accurate Warehouse Inventory

Most small warehouses run on a Warehouse Management System (WMS) like Fishbowl or a NetSuite module. These systems are fundamentally databases that trust human input. A worker scans a pallet into location A-01 but physically places it in A-02. The WMS records the pallet in A-01 and will report it there until a human manually corrects it. The system has no independent way to verify the physical location against the digital record.

In practice, this creates a constant drift between what the WMS reports and what is actually on the shelf. Consider a 15-person e-commerce fulfillment center. A picker is sent to bin C-04 for three units of SKU 123. They arrive to find the bin empty. The WMS says three units are there. Now, work stops. The picker spends 10 minutes searching nearby bins, finds the missing items in C-05, and completes the pick. This single error introduced 10 minutes of non-productive time, delayed an order, and still requires a manager to go back and fix the WMS record.

The structural problem is that a WMS is a system of record, not a system of observation. Its architecture is designed to process transactions (scan in, scan out) but not to perceive the state of the physical world. RF guns and barcode scanners make the transactions faster, but they do not solve the core issue. A worker can still scan the wrong item, enter the wrong quantity, or place a correctly scanned item in the wrong location. These tools cannot provide the feedback loop needed to keep the digital twin of the warehouse accurate.

We would start with a process and data audit. The first step is to understand your current workflow, your WMS's capabilities, and the most common sources of error. We would analyze your WMS data exports to understand your SKU velocity and inventory patterns. This discovery phase results in a clear architectural plan showing how a new AI validation layer would connect to your existing systems without disrupting operations.

The technical approach would involve a Python service, deployed on AWS Lambda for efficiency, that periodically pulls data from your WMS API. To validate this data, we would use computer vision. A worker could use a mobile app to take a picture of a bin. That image is sent to an endpoint powered by the Claude API's vision model, which identifies SKUs and counts them. The Python service then compares Claude's count to the WMS record for that bin. Discrepancies are logged in a Supabase database. The analysis of a single bin image typically takes less than 3 seconds.

The deliverable is a simple dashboard, hosted on Vercel, that presents a prioritized list of discrepancies for a manager to review. Instead of running blind cycle counts, your team would start each day with an actionable list of specific bins that need verification. This reduces manual count time by over 75%. You would receive the full source code for the Python service and the dashboard, along with a runbook for maintenance. The entire build, from discovery to deployment, is typically a 4-week process.