Production-scale ranking and recall models

The ranking and recall layers of an e-commerce recommendation stack. Deep-learning models that decide which products to retrieve and in what order to show them.

At hundreds of millions of users and billions of products, ordering matters more than catalog size. These models pick which products each shopper sees, in what order, and serve them under tight latency budgets.

Designed and trained deep-learning ranking and retrieval models on ByteDance's distributed inference infrastructure. The pipeline included offline feature stores, large-scale training, online serving with strict latency budgets, and shadow-then-A/B rollouts. Drove +93% GMV, +96% orders, and +39% CTR over the prior generation.

Co-built as part of the Algorithm team at ByteDance.