IBS Software cuts cargo NER costs 14x with Bedrock distillation
IBS Software deployed a bilingual named entity recognition system for cargo logistics using Amazon Bedrock's model distillation, extracting 23 entity types from English and Japanese email messages. The system distilled knowledge from Amazon Nova Pro into the lighter Nova Lite model, achieving 95.085 percent F1-Score accuracy while cutting operational costs by 14x. The solution processes thousands of cargo emails daily in real time, replacing manual intervention that previously slowed operations.
TL;DR
- IBS Software built a bilingual NER system extracting 23 entity types (AWB numbers, flight details, weights, delivery instructions) from cargo logistics emails in English and Japanese
- Used Amazon Bedrock model distillation to compress Nova Pro into Nova Lite, achieving 95.085% F1-Score accuracy with 14x cost reduction
- Team of 9 researchers and engineers completed the project in 4 months, annotating 500 bilingual emails (350 English, 150 Japanese) and training the student model over 70 steps
- System now processes thousands of cargo emails daily in real time, eliminating manual intervention bottlenecks
Why It Matters
Model distillation is emerging as a practical path to deploy AI systems at scale without prohibitive inference costs. This case demonstrates that smaller, specialized models can match larger ones on domain-specific tasks when properly trained, making enterprise AI deployment more economically viable for organizations processing high-volume multilingual data.
Business Impact
Cargo logistics relies on rapid, accurate data extraction from unstructured email. Manual intervention creates operational delays and errors. By automating entity extraction across two languages with 95% accuracy at 14x lower cost than alternatives, IBS Software reduced processing friction and improved throughput without sacrificing quality or requiring expensive infrastructure.
Key Implications
- Model distillation can deliver production-grade accuracy on specialized tasks while significantly reducing inference costs, making it viable for high-volume operational workflows
- Bilingual and multilingual NER is achievable with managed distillation tools, lowering the barrier for companies serving global supply chains
- Domain-specific annotation (500 emails) combined with knowledge distillation can outperform generic open-source frameworks, suggesting a shift toward purpose-built AI solutions over general-purpose tools
What to Watch
Monitor whether other logistics and supply chain companies adopt similar distillation approaches for multilingual document processing. Watch for adoption patterns across industries handling high-volume unstructured data in multiple languages, and track whether managed distillation becomes standard practice for cost-sensitive enterprise deployments.
Subscribe to the newsletter
The latest stories and analysis, delivered to your inbox.
Free. No spam. Unsubscribe any time.
