How Harmonic Security improved their data-leakage detection system with low-latency fine-tuned models using Amazon SageMaker, Amazon Bedrock, and Amazon Nova Pro

This article details how Harmonic Security optimized their data-leakage detection system using AWS services to achieve sub-500 millisecond latency while maintaining accuracy.

• Reduced detection latency from 1-2 seconds to under 500ms (76% improvement at median)
• Fine-tuned ModernBERT models (149M-395M parameters) instead of 8B parameter baseline
• Generated synthetic training data using Meta Llama 3.3 70B and Amazon Nova Pro
• Implemented binary and multi-label classification approaches for sensitive data detection
• Used Optuna for hyperparameter optimization across learning rate, dropout, batch size
• Deployed on SageMaker with ml.g5.4xlarge GPU instances and auto-scaling (1-5 instances)
• Achieved 48%-640% additional throughput capacity with auto-scaling enabled
• Binary classification: +1.56% accuracy, +2.26% F1 score improvement
• Solution now available on AWS Marketplace for enterprise deployment

Harmonic Security successfully optimized their AI governance platform by combining fine-tuned compact models with AWS infrastructure, delivering faster data protection without sacrificing accuracy or user experience.