Fiftify: AI demand forecasting and auto-replenishment

Fiftify's customers are multichannel sellers on Shopify, Amazon, eBay and retail, and most of them planned replenishment by hand: export stock to a spreadsheet, glance at last month's sales, place the order on instinct. The result cut both ways: cash locked up in slow movers and, at the same time, stockouts on bestsellers right at peak season.

They needed a model that accounts for seasonality, promotions and uneven supplier lead times. More importantly, it had to live inside the daily WMS workflow rather than in a BI dashboard nobody opens.

We built a forecasting service in Python (LightGBM) on top of Fiftify's sales and stock-movement history. The model trains per SKU and location and factors in seasonality, day of week, the promo calendar, sales channel and stockout windows, so a period of zero stock isn't read as falling demand.

It returns an 8-week forecast, safety stock and a reorder point for every item. Fiftify turns those into draft purchase orders and the buyer simply approves them.

The service runs standalone on FastAPI with a Redis queue and nightly retraining in Docker, talking to the WMS over an internal API. Accuracy is tracked on WAPE and benchmarked against manual planning on a control group of SKUs.