We consider the problem of “hyper-localizing” product assortments at a fashion retailer — that is, customizing the offerings to the particular preferences of customers visiting the store, so that customers can easily find the products that fit their tastes and purchase more. To make this decision, the firm must accurately predict the demand for each style at each store — a challenging task because of large variety and the small number of purchases per customer. To address this challenge, we propose: (a) a nonparametric choice modeling technique that uses purchase transactions tagged by customer IDs to build distributions over preference lists of products that are personalized to each customer and (b) an optimization framework that uses the predictions from our choice models to optimally allocate merchandise to different stores subject to inventory and dollar budget constraints. We implemented our methods at a large US fashion retailer with about $3B in annual revenue and approximately 300 stores. In a controlled experiment, our methods resulted in additional 7% revenue growth (approx. $200M profit impact) over the current method. We present the implementation details and the specific challenges (both technical and managerial) posed for assortment planning by fashion retail and the ways we addressed them.