Claim processing pipeline
Optimized health insurance claims processing using fine-tuned AI models and PaddleOCR for invoice extraction and policy exclusions, improving accuracy from 60% to 80% while cutting costs by 50%.
Business Challenge
A health insurance company struggled to accurately process medical invoices, hospital bills, handwritten documents, and poor-quality scans. The key challenge was extracting invoice line items, calculating the correct claimable amount, and identifying expenses related to diseases or treatments excluded from policy coverage.
Solution Delivered
Redesigned the health insurance document processing pipeline by replacing LLM-based extraction with a fine-tuned 4B parameter model specialized in forms and invoices. Migrated OCR from Tesseract to PaddleOCR to improve text extraction from handwritten documents and poor-quality scans. Additionally, fine-tuned a dedicated disease extraction model to replace OpenAI API calls, enabling more accurate identification of diseases and policy exclusions.
Outcomes Achieved
Improved overall document processing accuracy from 60% to 80% while reducing processing costs by 50%. Enabled more reliable extraction of medical invoices, accurate calculation of claimable amounts, and better identification of policy-excluded diseases and treatments, reducing manual intervention and improving claim processing efficiency.
Measurable Business Outcome
Accuracy improvement by 20% and cost reduction by 30%