Supply Chain Optimization System

01The challenge:

Ordering a quarter ahead from last year's spreadsheet

GlobalTrade Co imports commercial kitchen equipment and spare parts from factories in Vietnam, China and Mexico, and stocks four US warehouses. Sea freight takes ten to fourteen weeks, so buyers commit to orders a quarter ahead. They planned from a spreadsheet of last year's sales, adjusted by instinct, one warehouse at a time.

Slow items piled up in Atlanta while New Jersey ran out of the same fryer parts. The fix was air freight at six times the sea rate, plus carrying costs on stock nobody was buying. The owner saw the damage in the margins every month but could not trace it to particular orders.

02What we built:

A demand forecast per SKU and warehouse, and orders the buyer signs

Python jobs pull three years of order history, open purchase orders and supplier lead times from the ERP into PostgreSQL every Sunday night. A TensorFlow model forecasts weekly demand for 1,800 SKUs at each warehouse, 26 weeks out, using seasonality, dealer promotions and price changes, and reports its own uncertainty for every line.

A reorder engine turns the forecast into safety stock and order quantities, allowing for each supplier's lead-time spread and filling containers before splitting them. Buyers see the recommendations through a FastAPI service on Monday morning, approve or change them, and give a reason for any override. Everything runs in Docker on GlobalTrade's own server.

03What changed:

Costs down a quarter in six months, from buying closer to demand

Combined inventory and freight cost fell 25% in the first six months, mostly from lower holding costs on slow stock and far fewer air shipments. Stockouts on the 200 top-selling SKUs dropped 38%, and forecast error at the warehouse level fell from 34% to 19%.

Buyers still make the call, and they override about one line in ten, usually for a large dealer order the model cannot see coming. GlobalTrade's purchasing lead owns the approval screen and the override rules. We retrain the model on the first of each month and send the owner a one-page accuracy report.