IBS Software cuts cargo AI inference costs by 14× with Amazon Bedrock distillation

Key takeaway

• IBS Software deployed a bilingual cargo logistics system using Amazon Bedrock's model distillation to compress a teacher model into a smaller student model, achieving 95.085 percent accuracy on named-entity recognition while cutting inference costs by 14×.

• The approach retained 98 percent of the teacher model's performance and now processes cargo emails in real time, demonstrating that managed distillation can balance cost and accuracy for production multilingual systems.

3 Key Points

FAQ

How much accuracy did the distilled model retain compared to the teacher model?

The distilled Nova Lite student model achieved 95.085 percent F1-Score overall, retaining 98 percent of the teacher model's performance. The teacher Nova Pro model scored 97.0 percent overall, with English at 97.8 percent and Japanese at 96.2 percent.

What was the training process and timeline?

IBS Software's team of nine researchers and engineers spent approximately 4 months. The student model was trained for 4 epochs over 70 steps, reducing loss from 0.05 to 0.008, using 500 manually annotated cargo email messages (350 English, 150 Japanese) with 23 entity types.

What entities does the system extract from cargo emails?

The system identifies 23 entity types including air waybill (AWB) numbers, flight numbers and routes, weights (gross, chargeable, dimensional), dimensions and volume, commodity descriptions, shipper and consignee information, special handling codes, and delivery instructions.