IBS Software cuts bilingual NER costs 14x using Amazon Bedrock distillation

IBS Software's cargo logistics system handles thousands of bilingual (English and Japanese) email messages daily, requiring extraction of 23 entity types such as air waybill numbers, flight details, weights, and delivery instructions. Initial attempts with open-source frameworks like PyTorch and TextBrewer failed due to complexity in configuring distillation pipelines for bilingual data, lack of managed infrastructure, and difficulty tuning token-level hyperparameters. A team of nine researchers spent two months struggling with these approaches before pivoting to Amazon Bedrock Model Distillation.

Using Amazon Nova Pro as the teacher model and Amazon Nova Lite as the student model, they achieved 95.085% F1-score accuracy while reducing operational costs by 14x. The managed training infrastructure handled automatic hyperparameter optimization and native token-level distillation, enabling deployment in just two months. The production system now processes .eml files in real time, eliminating manual intervention and providing significant cost savings at scale. Key achievements include training over 4 epochs with 70 steps, reducing loss from 0.05 to 0.008, and validating on 500 annotated emails.