Amazon, in partnership with the Indian Institute of Technology, Kharagpur, has unveiled the Rate Card Transformer (RCT) framework for predicting shipping costs on day 0.

The RCT model is built on the Transformer framework and uses multiple encoder blocks to analyze various package features such as dimensions, route, service, item, and charge.

These features are embedded into tokens and processed through Transformer encoder layers, allowing the model to learn intricate interactions and context-aware representations.

The RCT model is trained to minimize L1 loss and captures dynamic rate card information, outperforming traditional models like GBDT and neural networks.

The model's scalability and efficiency in predicting shipping costs highlight the effectiveness of transformer-based models in this domain.