Sun Finance and the AWS Generative AI Innovation Center rebuilt the company's identity verification pipeline, improving ID extraction accuracy from 79.7% to 90.8% on 585 ID images, cutting per-document costs by 91% and reducing processing time from up to 20 hours to under 5 seconds. The solution went live in production on January 22, 2026.
"What initially felt like an ambitious — almost unrealistic — target has been transformed into a secure, production-ready solution delivering measurable gains in accuracy, speed, and cost efficiency," said Agris Vaselāns, Group CRO of Sun Finance.
Sun Finance, a Latvian fintech founded in 2017, operates as a technology-first online lending marketplace across nine countries. It processes a new loan request every 0.63 seconds and delivers more than 4 million evaluations monthly. In one high-volume market with 80,000 monthly applications for microloans, approximately 60% required manual operator review. About 80% of those cases stemmed from mismatches between extracted information and customer-entered data, and 60% of those mismatches were OCR errors; the remaining 20% of manual interventions related to fraud detection flags. The per-document cost and about 3 full-time equivalents on manual verification in that region blocked expansion into lower-value microloans.
The engagement ran from kickoff on August 26, 2025 to a final presentation on October 9, 2025, with technical handover on November 14, 2025. Sun Finance then took 35 business days to reach production, including a 14-day production freeze from December 18 to January 7.
The serverless solution exposes two routes through Amazon API Gateway. On /extract-id, AWS Lambda sends the image to Amazon Textract for primary OCR, falls back to Amazon Rekognition when confidence is low, and passes the text to Claude Sonnet 4 on Amazon Bedrock, which structures it into standardized JSON fields, followed by validation rules. On /detect-fraud, an AWS Step Functions workflow runs two checks in parallel: background similarity, using Amazon Rekognition face masking, Amazon Titan Multimodal Embeddings and Amazon S3 Vectors, and visual pattern detection by Claude Sonnet 4 for screen photos and digital manipulation. Authentication uses Amazon Cognito with AWS SigV4, alongside AWS WAF and AWS KMS, with Terraform-defined infrastructure.
Claude Sonnet 4 alone reached 61.8% accuracy, Amazon Textract plus Claude structuring reached 85%, and the multi-tier version reached 90.8%. The end-to-end fraud detection pipeline achieved 81% accuracy, 59% recall and 83% specificity. Manual intervention is projected to drop from 60% to 30% of applications and staffing from about 3 FTEs to about 1.
Details