ProductionEvidence: Medium60/100

Sun Finance automates ID extraction and fraud detection with generative AI on AWS

Use case typeFraud detectionUpdated Jun 13, 2026

Sun Finance, a fintech online lending marketplace operating in nine countries, faced challenges with high manual workload for identity document verification and fraud detection due to OCR errors and complex document types across multiple languages. About 60% of loan applications required manual review, resulting in high costs and slow processing times up to 20 hours. They partnered with the AWS Generative AI Innovation Center to build an AI-powered identity verification pipeline and a serverless fraud detection system using Amazon Bedrock (Anthropic Claude Sonnet 4, Amazon Titan Multimodal Embeddings), Amazon Textract, Amazon Rekognition, AWS Step Functions, Amazon API Gateway, AWS Lambda, and Amazon S3 Vectors. The solution architecture uses multi-tier OCR extraction combined with LLM structuring and vector similarity search for fraud pattern detection. Amazon Textract handles primary OCR, Amazon Rekognition is the fallback for low-confidence OCR, and Amazon Bedrock structures extracted text into JSON. Fraud detection combines visual pattern recognition and background similarity analysis using vector search against known fraud patterns. The system increased extraction accuracy from 79.7% to 90.8%, cut per-document costs by 91%, reduced processing time from 20 hours to under 5 seconds, halved manual review workload, and enabled cost-effective scaling to serve lower-value microloan markets.

Organization
Sun Finance
Industry
Finance
Location
Latvia
Published
April 2026

Reported outcomes

−91%

costCost savings

79.7-90.8%accuracy20 hourstime5 secondstime+81%accuracy+59%accuracy

Strategic outcomes

Speed & agilityReduced document processing to secondsCost efficiencyHalved manual review workloadMarket & geographic expansionExpanded into lower-value loan segmentsNew product / capabilityBuilt AI-powered fraud detection system

Catalog median for cost savings deployments: −40% across 177 reported metrics. Compare benchmarks →

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

Normalized claim

Accuracy: 79.7-90.8% increase

AWS Machine Learning BlogApr 30, 2026Blog postInferred claimMedium evidence strength

The new solution improved ID extraction accuracy from 79.7% to 90.8%.

Normalized claim

Time: 20 hours decrease

AWS Machine Learning BlogApr 30, 2026Blog postInferred claimMedium evidence strength

Processing time per document dropped from up to 20 hours to under 5 seconds.

Normalized claim

Time: 5 seconds decrease

AWS Machine Learning BlogApr 30, 2026Blog postInferred claimMedium evidence strength

Processing time per document dropped from up to 20 hours to under 5 seconds.

Normalized claim

Cost: 91% decrease

AWS Machine Learning BlogApr 30, 2026Blog postInferred claimMedium evidence strength

Per-document costs fell by 91%, allowing expansion into lower-value loan segments.

Normalized claim

Accuracy: 81% increase

AWS Machine Learning BlogApr 30, 2026Blog postInferred claimMedium evidence strength

The fraud detection pipeline achieved 81% accuracy with 59% recall, and the system improves as more fraud cases are added to the reference database.

Normalized claim

Accuracy: 59% increase

AWS Machine Learning BlogApr 30, 2026Blog postInferred claimMedium evidence strength

The fraud detection pipeline achieved 81% accuracy with 59% recall, and the system improves as more fraud cases are added to the reference database.

Normalized claim

Quantified impact: 40% decrease

AWS Machine Learning BlogApr 30, 2026Blog postInferred claimMedium evidence strength

The serverless architecture enabled rapid updates and parallel execution to reduce fraud detection latency by 40%.

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks
Customer
Sun Finance
Provider
AWS
Maturity
Production

Manual review workload was halved, reducing operational costs and enabling economic viability for microloan markets

Customer identity supportedSource describes one deploymentMaturity supported

Primary read

Use case focus

Showing 2 of 2

  • 1Identity Verification
  • 2Fraud Detection
  • Sun Finance had a high manual workload to verify identity documents and detect fraud, caused by OCR errors and the complexity of document types in multiple languages.
  • OCR errors alone caused around 60% of loan applications to require manual review, which was costly and slow, with processing times up to 20 hours.
  • Fraud detection was also complex, with fraudsters submitting similar images with distinctive patterns to bypass controls, requiring time-intensive manual review.
  • Sun Finance collaborated with the AWS Generative AI Innovation Center to develop an AI-powered ID extraction pipeline and a serverless fraud detection system using AWS AI services.
  • The ID extraction pipeline used Amazon Textract for primary OCR, Amazon Rekognition as a fallback OCR, and Amazon Bedrock running Anthropic Claude Sonnet 4 to structure extracted text.
  • The fraud detection pipeline ran two parallel checks orchestrated by AWS Step Functions: visual pattern detection via Claude Sonnet 4 and background similarity analysis using Amazon Titan Multimodal Embeddings and Amazon S3 Vectors vector search against known fraud patterns.
  • The architecture was fully serverless using AWS Lambda and Amazon API Gateway, allowing rapid iteration and deployment with Terraform.
Technologies
Amazon BedrockAmazon TextractAmazon RekognitionAWS Step FunctionsAmazon API GatewayAWS LambdaAmazon S3 VectorsAnthropic Claude Sonnet 4Amazon Titan
  • The new solution improved ID extraction accuracy from 79.7% to 90.8%.
  • Processing time per document dropped from up to 20 hours to under 5 seconds.
  • Manual review workload was halved, reducing operational costs and enabling economic viability for microloan markets.
  • Per-document costs fell by 91%, allowing expansion into lower-value loan segments.
  • The fraud detection pipeline achieved 81% accuracy with 59% recall, and the system improves as more fraud cases are added to the reference database.
  • The serverless architecture enabled rapid updates and parallel execution to reduce fraud detection latency by 40%.
Architecture

The solution uses a serverless architecture with two API routes exposed through Amazon API Gateway. The ID extraction route uses AWS Lambda to process images via Amazon Textract for primary OCR and Amazon Rekognition as fallback OCR. Extracted text is structured by Anthropic Claude Sonnet 4 running on Amazon Bedrock. The fraud detection route triggers AWS Step Functions workflows that run visual pattern detection and background similarity analysis in parallel, with results combined by a Lambda risk assessment function. Confirmed fraud images are ingested, processed, vectorized, and stored in Amazon S3 Vectors for growing the fraud pattern database.

Implementation partners1
AWS Generative AI Innovation Center
Sources & evidence3
Evidence: Medium60/100Evidence strength
  • Customer explicitly identified
  • Deployment status explicitly supported
  • Quantified outcome available
  • Technical implementation details available
  • Multiple corroborating sources available
Type: Blog PostPublished: Apr 30, 2026Publisher: AWS Machine Learning BlogEvidence: VendorConfidence: Medium
Primary sourceSource 2Source 3

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

Explore related AI use cases
Customer evidence
More from Sun Finance
Provider evidence
AWS provider pageAWS ranked cases
Market context
Finance AI use casesFinance industry insightsLatvia country insights
AI patterns
RAGVision
Browse the catalog
All AI use casesAI use-case typesFraud detection use cases

Was this useful?

Community

Comments

No published comments yet.