Custom AI to Optimize Your Warehouse Picking and Packing Workflows

Many logistics companies rely on the built-in features of their WMS, like NetSuite WMS or Fishbowl. These systems are excellent for inventory accounting but offer only rudimentary picking logic, such as simple zone-based or wave picking. The logic is static. It cannot dynamically re-route a picker around a temporarily congested aisle or prioritize a last-minute high-priority order by re-calculating an entire batch.

Consider a 15-person warehouse team for an e-commerce distributor with 5,000 SKUs. A picker receives a list of ten single-item orders from the WMS, sorted alphabetically by customer. The first item is in Aisle 1, the second in Aisle 12, the third back in Aisle 2. The picker spends an hour walking 3 miles, crisscrossing the floor to complete a batch that a smarter route would fulfill in a 1-mile path. Barcode scanners confirm the correct item is picked, but they do nothing to make the process more efficient.

The structural problem is that WMS platforms are systems of record, not operational optimization engines. Their architecture is designed for transactional integrity, ensuring inventory counts are accurate. They are not built to run complex graph-based routing algorithms on real-time spatial data. You cannot inject a custom optimization model into their compiled, closed-source workflow engine. This forces teams into inefficient manual processes that directly limit order throughput.

The engagement would begin with a warehouse audit. We would map your physical layout, document SKU-to-bin locations, and analyze six months of historical order data from your WMS. This process identifies common picking patterns, logical order groupings, and the specific API endpoints or data export methods available from your current systems. You receive a report detailing the potential efficiency gains before any development starts.

The core of the system would be a Python service built with FastAPI, deployed on AWS Lambda for event-driven processing that keeps hosting costs low (typically under $50/month). When a new batch of orders is ready, your WMS sends the data to a secure endpoint. The service uses the `networkx` library to model the warehouse as a graph and an A* search algorithm to compute the shortest picking path for the batch. We have built Claude API pipelines for parsing unstructured financial documents, and the same pattern applies to extracting special handling instructions from order notes.

The final deliverable is a system that sends optimized pick lists directly back to your team's existing hardware, whether handheld scanners or printers, via an API call to your WMS or a simple web interface. Each list is sequenced to minimize travel. You receive the full Python source code in your GitHub repository, a runbook explaining how to update the warehouse map, and a Supabase dashboard to track key metrics like picks per hour and average time per order.