What challenges did IBS Software face before AI implementation?

Process thousands of bilingual (English/Japanese) cargo logistics .eml messages daily. Extract 23 entity types including AWB numbers, flight details, weights, and delivery instructions with high accuracy. Keep inference costs low and achieve low-latency real-time processing.

What was the impact of the AI solution?

The solution achieved 95.085% F1-score accuracy. It processed messages in under 2 seconds. It reduced operational inference costs by 14x while retaining about 98% of the teacher model performance.

Source

IBS Software builds bilingual NER for cargo logistics emails using Amazon Bedrock managed distillation

Use case typeDocument processing automationUpdated Jun 30, 2026

IBS Software's cargo system processes thousands of bilingual cargo logistics email messages daily, extracting critical information such as air waybill numbers, flight details, weights, and delivery instructions in English and Japanese. The team built a production-ready bilingual named entity recognition solution to identify 23 entity types across the two languages while keeping inference cost low and supporting real-time processing.

Organization: IBS Software
Industry: Logistics
Location: United States
Published: June 2026

Reported outcomes

98%

teacher performance retainedOther quantified impact

+95.1%F1-score accuracy14xoperational inference cost2 secondsprocessing latency

Strategic outcomes

Speed & agilityEnabled real-time cargo email processingCost efficiencyReduced inference spend while preserving accuracy

Primary read

Use case focus

Showing 2 of 2

1Document processing automation
2AI model training

Process thousands of bilingual (English/Japanese) cargo logistics .eml messages daily.
Extract 23 entity types including AWB numbers, flight details, weights, and delivery instructions with high accuracy.
Keep inference costs low and achieve low-latency real-time processing.

IBS Software annotated 500 bilingual email messages, with 350 English and 150 Japanese examples, for 23 entity types.
The team used Amazon Bedrock managed distillation to transfer knowledge from Amazon Nova Pro to Amazon Nova Lite with token-level distillation.
They deployed a pipeline where Amazon S3 receives .eml files, AWS Lambda extracts content, Amazon Bedrock runs the distilled model, and structured JSON results are stored in Amazon DynamoDB with confidence filtering and validation rules.

Technologies

Amazon Bedrock Amazon Nova Pro Amazon Nova Lite Amazon S3 AWS Lambda Amazon DynamoDB Bedrock Runtime

The solution achieved 95.085% F1-score accuracy.
It processed messages in under 2 seconds.
It reduced operational inference costs by 14x while retaining about 98% of the teacher model performance.

Architecture

Cargo email messages arrive as .eml files in Amazon S3. AWS Lambda extracts email body and metadata. Amazon Bedrock processes text with a distilled Nova Lite model trained via managed distillation from Nova Pro. The model returns 23 entity types with confidence scores, then validation rules and confidence filtering are applied before structured JSON is stored in Amazon DynamoDB.

Sources & evidence1

Groundedness: 5/5Type: Blog PostPublished: Jun 30, 2026Publisher: AWSEvidence: VendorConfidence: High

Primary source

AI-generated summary. Verify important details with the linked sources before relying on this case.

Explore related AI use cases

Customer evidence