What was the impact of the AI solution?

The fine-tuned Nova Micro model achieved up to 94.77% extraction accuracy. Accuracy improved by up to 16.6 percentage points over the baseline. Inference latency was reduced by more than 30 percent. Inference costs were cut by about 50 percent. Parcel Perform moved the solution into production.

Source

ProductionEvidence: Medium50/100

Parcel Perform fine-tunes Amazon Nova for ecommerce email entity extraction

Use case typeDocument processing automationUpdated Jun 30, 2026

Parcel Perform, an AI delivery experience platform for ecommerce businesses, needed to extract structured information from diverse email formats, including HTML-heavy messages with JavaScript elements. The company worked with the AWS Generative AI Innovation Center to fine-tune Amazon Nova Micro and Nova Lite models for accurate entity extraction from ecommerce emails.

Organization: Parcel Perform
Industry: Logistics
Location: Singapore
Published: June 2026

Reported outcomes

+94.8%

extraction accuracyQuality & accuracy

+16.6%accuracy improvement over baseline−30%inference latency−50%inference cost

Strategic outcomes

Cost efficiencyMoved the solution into productionOther strategic outcomeReduced hallucinations in extraction

Why do we believe this?Outcome claims, sources, and evidence checks

Normalized claim

Extraction accuracy: 94.8% increase

AWS Machine Learning BlogJun 30, 2026Blog postExplicit claimMedium evidence strength

the fine-tuned Nova Micro models achieved up to 94.77% extraction accuracy

Normalized claim

Accuracy improvement over baseline: 16.6% increase

AWS Machine Learning BlogJun 30, 2026Blog postExplicit claimMedium evidence strength

an improvement of up to 16.6 percentage points over the baseline

Normalized claim

Inference latency: 30% decrease

AWS Machine Learning BlogJun 30, 2026Blog postExplicit claimMedium evidence strength

reduced inference latency by more than 30 percent

Normalized claim

Inference cost: 50% decrease

AWS Machine Learning BlogJun 30, 2026Blog postExplicit claimMedium evidence strength

halved costs compared with Parcel Perform’s previous model

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: Parcel Perform
Provider: AWS
Maturity: Production
Linked source: AWS Machine Learning Blog

The team prepared training data in the Amazon Bedrock conversation format, uploaded it to Amazon S3, and deployed the tuned model in Amazon Bedrock for on-demand inference

Customer identity supportedSource describes one deploymentMaturity supported

Primary read

Use case focus

Showing 2 of 2

1Document processing automation
2Content processing automation

Extract accurate structured entities from high-volume, diverse e-commerce emails while reducing hallucinations and token and inference costs.

Parcel Perform used Amazon SageMaker AI to supervised fine-tune Amazon Nova Micro and Nova Lite models with Parameter-Efficient Fine-Tuning using LoRA.
The team prepared training data in the Amazon Bedrock conversation format, uploaded it to Amazon S3, and deployed the tuned model in Amazon Bedrock for on-demand inference.
The fine-tuned model was moved into production to support parcel and tracking data extraction in ecommerce logistics operations.

Technologies

Amazon SageMaker AI Amazon Bedrock Amazon S3 AWS IAM Amazon Nova

The fine-tuned Nova Micro model achieved up to 94.77% extraction accuracy.
Accuracy improved by up to 16.6 percentage points over the baseline.
Inference latency was reduced by more than 30 percent.
Inference costs were cut by about 50 percent.
Parcel Perform moved the solution into production.

Architecture

Training data was prepared in Amazon Bedrock conversation format, uploaded to Amazon S3, and used in an Amazon SageMaker AI supervised fine-tuning job with LoRA/PEFT on Amazon Nova Micro and Nova Lite. The custom model was then imported into Amazon Bedrock for on-demand inference.

Sources & evidence1

Evidence: Medium50/100Evidence strength