Sun Finance Automates ID Extraction and Fraud Detection

According to an AWS blog post co-authored with Sun Finance staff, Sun Finance built an identity verification (IDV) pipeline that combines `Amazon Textract`, `Amazon Rekognition`, and `Amazon Bedrock`. The post reports the deployment improved extraction accuracy from 79.7% to 90.8%, reduced per-document costs by 91%, and shortened processing time from as long as 20 hours to under 5 seconds. The blog states the AWS Generative AI Innovation Center engagement ran 32 days from kickoff through final presentation, followed by a 26-day technical handover and a 35-business-day push to production, with the system going live on January 22, 2026.

According to the AWS post, the solution pairs specialized OCR for structured fields with LLM-based structuring to extract and normalize data, and it implements serverless fraud detection using vector similarity search to surface near-duplicate or suspicious identity artifacts. The blog frames the architecture as serverless and cloud-native and highlights that the hybrid approach outperformed using OCR or LLMs alone for Sun Finance's dataset and regional document complexity.

Industry-pattern observations: Combining high-accuracy OCR with LLM-based post-processing is a common pattern for large-scale document automation because OCR preserves layout and deterministic extraction while LLMs help normalize, infer, and map noisy text to canonical records. Similarly, using vector similarity search as part of fraud controls is increasingly common for detecting near-duplicates and identity reuse at scale.

Editorial analysis: The reported metric improvements-double-digit accuracy gains, multi-hour to sub-second latency reductions, and a 91% per-document cost cut-illustrate why lenders and fintechs are adopting generative-AI-assisted pipelines for IDV. For practitioners, the case underscores practical priorities: orchestration between specialized extractors and LLMs, production-grade latency and cost optimization, and integration of similarity search for fraud signals.

For practitioners: monitor manual-review rate, end-to-end false positive and false negative rates, model drift on regional document types, latency and cost under real traffic, and auditability for compliance. Observers should also watch how vendors surface tooling for grounding LLM outputs to deterministic OCR fields and how vector search scales with historical identity stores.