Blab AI
IBS Software Cuts AI Costs 14-Fold with Amazon Bedrock Distilled Model for Bilingual Cargo Email Processing
IBS Software has slashed the cost of its real‑time cargo email processor by a factor of 14, all while maintaining near‑perfect accuracy. The cloud‑based air‑cargo firm announced that it used Amazon Bedrock’s model‑distillation feature to shrink the larger Amazon Nova Pro foundation model into a leaner Nova Lite version. The new model still pulls 23 different entity types from English and Japanese cargo logistics emails with a 95.085 % F1‑score.
The engine behind the savings is part of IBS Software’s iCargo platform, which receives thousands of bilingual cargo emails each day. Those messages carry critical data—air waybill numbers, flight details, weight and volume measurements, commodity descriptions, and delivery instructions—whose accurate extraction is essential for downstream logistics and billing.
Building the bilingual named‑entity recognition (NER) model took a dedicated team of nine researchers and engineers four months. In month 1 they assembled a dataset of 500 real‑world emails—350 in English and 150 in Japanese—annotated by domain experts for 23 entity categories. The three‑week annotation effort produced high‑quality training data for both languages.
Month 2 saw the team experiment with open‑source frameworks such as PyTorch and the TextBrewer library. Those attempts ran into roadblocks: the frameworks lacked built‑in support for token‑level distillation on bilingual data, required manual infrastructure for training and deployment, and made hyperparameter tuning cumbersome. In addition, the pipelines did not integrate cleanly with IBS Software’s existing email ingestion workflow.
In month 3 the team switched to Amazon Bedrock’s managed distillation workflow. Using Nova Pro as the teacher model and Nova Lite as the student, they configured custom hyperparameters and trained the student for four epochs over 70 steps. The training process reduced the loss from 0.05 to 0.008, indicating effective knowledge transfer. After distillation, the Nova Lite model achieved a 95.085 % overall F1‑score, with 96.535 % on English and 93.635 % on Japanese. The teacher model scored 97.0 % overall, so the distilled model retained 98 % of the teacher’s performance.
The deployment architecture relies on AWS services that require minimal custom code. Incoming .eml files are stored in Amazon S3, where an AWS Lambda function extracts the email body and metadata. The Lambda function then calls a Bedrock inference endpoint that runs the distilled Nova Lite model. The model returns 23 entity types with confidence scores, which a post‑processing layer filters using rule‑based validation and confidence thresholds. The final structured JSON is written to Amazon DynamoDB. The entire pipeline processes an email in under two seconds, meeting the real‑time requirement of IBS Software’s cargo operations.
The cost savings are significant. The case study reports that the distilled Nova Lite model reduces inference costs by a factor of 14 compared to the base Nova Lite model, while still delivering a 10 % uplift in accuracy over the base model. This advantage is crucial for a high‑volume production workload that processes thousands of emails daily.
Looking ahead, IBS Software plans to expand the model’s coverage by adding synthetic Japanese examples to address remaining gaps in kanji‑heavy commodity descriptions. The company also intends to explore additional Bedrock foundation models for other logistics use cases.
In short, IBS Software’s experience shows that Amazon Bedrock’s managed distillation can simplify the development of bilingual NER systems for specialized domains. By leveraging a teacher–student workflow, the company achieved near‑teacher accuracy at a fraction of the cost and with a streamlined deployment pipeline.
The engine behind the savings is part of IBS Software’s iCargo platform, which receives thousands of bilingual cargo emails each day. Those messages carry critical data—air waybill numbers, flight details, weight and volume measurements, commodity descriptions, and delivery instructions—whose accurate extraction is essential for downstream logistics and billing.
Building the bilingual named‑entity recognition (NER) model took a dedicated team of nine researchers and engineers four months. In month 1 they assembled a dataset of 500 real‑world emails—350 in English and 150 in Japanese—annotated by domain experts for 23 entity categories. The three‑week annotation effort produced high‑quality training data for both languages.
Month 2 saw the team experiment with open‑source frameworks such as PyTorch and the TextBrewer library. Those attempts ran into roadblocks: the frameworks lacked built‑in support for token‑level distillation on bilingual data, required manual infrastructure for training and deployment, and made hyperparameter tuning cumbersome. In addition, the pipelines did not integrate cleanly with IBS Software’s existing email ingestion workflow.
In month 3 the team switched to Amazon Bedrock’s managed distillation workflow. Using Nova Pro as the teacher model and Nova Lite as the student, they configured custom hyperparameters and trained the student for four epochs over 70 steps. The training process reduced the loss from 0.05 to 0.008, indicating effective knowledge transfer. After distillation, the Nova Lite model achieved a 95.085 % overall F1‑score, with 96.535 % on English and 93.635 % on Japanese. The teacher model scored 97.0 % overall, so the distilled model retained 98 % of the teacher’s performance.
The deployment architecture relies on AWS services that require minimal custom code. Incoming .eml files are stored in Amazon S3, where an AWS Lambda function extracts the email body and metadata. The Lambda function then calls a Bedrock inference endpoint that runs the distilled Nova Lite model. The model returns 23 entity types with confidence scores, which a post‑processing layer filters using rule‑based validation and confidence thresholds. The final structured JSON is written to Amazon DynamoDB. The entire pipeline processes an email in under two seconds, meeting the real‑time requirement of IBS Software’s cargo operations.
The cost savings are significant. The case study reports that the distilled Nova Lite model reduces inference costs by a factor of 14 compared to the base Nova Lite model, while still delivering a 10 % uplift in accuracy over the base model. This advantage is crucial for a high‑volume production workload that processes thousands of emails daily.
Looking ahead, IBS Software plans to expand the model’s coverage by adding synthetic Japanese examples to address remaining gaps in kanji‑heavy commodity descriptions. The company also intends to explore additional Bedrock foundation models for other logistics use cases.
In short, IBS Software’s experience shows that Amazon Bedrock’s managed distillation can simplify the development of bilingual NER systems for specialized domains. By leveraging a teacher–student workflow, the company achieved near‑teacher accuracy at a fraction of the cost and with a streamlined deployment pipeline.
