Sun Finance: GenAI-Powered ID Verification and Fraud Detection

Title

Sun Finance: GenAI-Powered ID Verification and Fraud Detection

Content

Sun Finance, a Latvian online lender operating across nine countries, rebuilt its identity verification pipeline with the AWS Generative AI Innovation Center and went live January 22, 2026 — pushing extraction accuracy from 79.7% to 90.8% across 585 test ID images, cutting per-document cost by 91%, and reducing processing time from up to 20 hours to under 5 seconds. "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. The pressure: Sun Finance processes a new loan request every 0.63 seconds and delivers more than 4 million evaluations monthly. In its highest-volume microloan market, 80,000 monthly applications produced 60% manual review load — 80% OCR mismatch, 20% fraud flagging. Per-document cost plus ~3 FTEs of manual verification blocked expansion into lower-value microloan economies. The project spanned 107 business days across four milestones — AWS GenAI Innovation Center kickoff August 26, 2025, final presentation October 9 (32 days); 26 days technical handover to November 14; 35 business days to production including a 14-day holiday freeze (December 18 – January 7); live January 22, 2026. The architecture is fully serverless on AWS: Amazon API Gateway exposes `/extract-id` (Amazon Textract for primary OCR, Amazon Rekognition fallback, Amazon Bedrock with Claude Sonnet 4 for JSON structuring across 7 fields) and `/detect-fraud` (AWS Step Functions orchestrating two parallel checks — background similarity via Amazon Rekognition face-mask → Amazon Titan Multimodal Embeddings → Amazon S3 Vectors; and visual pattern detection via Claude Sonnet 4 for screen-photo bezels, scan lines, glare, and manipulation artifacts). Auth via Amazon Cognito + AWS SigV4, AWS WAF, AWS KMS at rest, TLS 1.2+ in transit, infrastructure in Terraform. Operational scale: the team iterated ID extraction through three approaches — Claude Sonnet 4 alone (61.8%, blocked by PII safety protocols), Textract + Claude structuring (85%, +11.6pp), then multi-tier OCR with validation rules (90.8%). Per-field gains: name 84.93%→87.72%, date of birth 81.25%→90.80%, document type 78.43%→96.40%, ID number 74.32%→89.40%. End-to-end fraud detection: 81% accuracy, 59% recall, 83% specificity. Visual embeddings beat text embeddings for background similarity (96/80/52 vs 91/27.8/21.7). Manual review projected to drop from 60% to 30% of applications; headcount for this market from ~3 FTEs to ~1 FTE. Next steps: extend visual detection to cartoons and AI-generated images; grow the fraud pattern database to lift background similarity recall; add EXIF metadata, device fingerprinting, and geolocation checks as new parallel Lambda branches; roll out language-specific prompt engineering across Southeast Asia, Africa, Latin America, and other European markets without code changes.