Amazon Nova fine-tuning boosts email extraction accuracy to 94.77%
Parcel Perfect cut costs by 50% and latency by 30% after fine-tuning Nova models.
Processing millions of email messages daily for ecommerce logistics requires accurate extraction of order numbers, tracking numbers, and other structured data from diverse email formats — from plain text to complex HTML with JavaScript. Common pitfalls include model hallucinations, confusion between similar fields, and high token costs. Parcel Perform, a global AI Delivery Experience Platform, partnered with the AWS Generative AI Innovation Center (GenAIIC) to fine-tune Amazon Nova models using supervised fine-tuning (SFT) with Parameter-Efficient Fine-Tuning (PEFT) via Low-Rank Adaptation (LoRA).
The fine-tuned Amazon Nova Micro model achieved 94.77% extraction accuracy on test data, a 16.6 percentage point improvement over the baseline. It also reduced inference latency by over 30% and cut costs by 50% compared to Parcel Perform’s previous model, matching or exceeding the fine-tuned Nova Lite at a lower price. The workflow uses SageMaker AI for customization, with training data in JSONL format stored in Amazon S3, and deploys the model on Amazon Bedrock for on-demand inference. This approach enables other enterprises to automate email data extraction with high accuracy, low latency, and significantly reduced operational costs.
- Fine-tuned Nova Micro achieved 94.77% accuracy, a 16.6 percentage point improvement over baseline.
- Inference latency reduced by over 30% and costs halved compared to Parcel Perform's previous model.
- Uses SFT with PEFT (LoRA) on SageMaker AI, deployed via Amazon Bedrock on-demand inference.
Why It Matters
Enterprises processing massive email volumes can now extract critical data with near-perfect accuracy at half the cost.