How Sun Finance and AWS sped up document checks and cut fraud risk

Sun Finance has moved identity verification to an automated mode using AWS services and generative AI. The company built a pipeline that extracts data from documents faster, processes applications more cheaply, and helps catch suspicious matches in near real time.

How the pipeline works

The solution splits the work across several tools. Amazon Textract handles OCR and pulls text from identity documents, Amazon Bedrock puts the result into the required structure and normalizes the fields, and Amazon Rekognition assists with visual checks and anti-fraud scenarios. This approach proved more practical than trying to solve everything with a single model: a specialized service does the extraction, while the LLM handles interpretation, format correction and assembly of the final structure for internal systems.

The architecture's speed matters separately. Sun Finance built a serverless IDV pipeline where processing is triggered by events and does not require permanent infrastructure sized for peak loads. This is especially useful for lending and fintech workflows, where document flow is uneven: at night it may be minimal, then spike sharply during loan-issuing hours. This setup makes costs easier to control, and response time no longer depends on a manual queue of operators, who previously could delay a check by many hours.

• OCR extracts raw text and fields from the document
• The LLM puts the data into a uniform structure and removes noise
• Rekognition adds visual signals for identity verification
• Vector search helps find suspiciously similar applications

For anti-fraud in this setup, what matters is not only reading the document but also being able to compare it with cases already processed. Vector search stores representations of applications and quickly finds close matches by image, text or a combination of features. If the system sees overly similar documents, repeating patterns or unusual proximity between different applications, it can send them for additional review. This gives a more flexible layer of protection than simple rigid rules.

Why it worked

The main takeaway of the case is that combining OCR and an LLM gives better results than using either component alone. According to Sun Finance, extraction accuracy rose from 79.7% to 90.8%. For such scenarios this is not a cosmetic improvement: every few percentage points of accuracy directly affect the volume of manual rework, the number of errors in application forms and the speed of the decision on an application. The LLM here does not replace specialized recognition but complements it: it understands the context of fields, aligns names, removes typical OCR artifacts and prepares data for further automation.

The economic effect is just as noticeable. Per AWS's description, the cost of processing one document fell by 91%, and processing time dropped from a maximum of 20 hours to under 5 seconds. For fintech this means several things at once: lower operating costs, fewer abandonments due to long waits, and a better chance of verifying the customer before they drop the application. In parallel, the anti-fraud system based on vector similarity can search for repeated or overly similar applications without heavy server-side plumbing. This makes verification not only faster but also more resistant to attempts to get around basic rules.

What this means

The Sun Finance case shows that generative AI in document workflows works best not as a universal replacement for the whole stack, but as a layer on top of already strong specialized services. For banks, microfinance lenders and other companies with KYC processes, the signal is clear: if OCR, structuring and anti-fraud are split across the right components, you can raise accuracy, cut response time to seconds and noticeably reduce the cost of verification at the same time.