Streamlining Mortgage Processing with AI OCR at P. a Mortgage Broking software in stealth development

Key takeaway: multiple trained custom document extractors will get you to higher F1 scores faster, and reduce your training costs.
Always start with training before developing the product.

Background and Context

• Company Profile: P. Stealth Mortgage Brokers, a part of a more extensive network, is an innovative mortgage broking firm with decades of experience in the industry. They leverage advanced technologies like Google Cloud Platform's Document AI to enhance their service offerings.
• Pre-Implementation Challenges: Before AI OCR implementation, P. Stealth Mortgage Brokers faced significant delays in the mortgage application process. The firm experienced a 5-10 day back-and-forth with clients to determine eligibility for regulated banking products. Manual transcription of PDF documents led to inaccuracies and inefficiencies.

Solution Implementation

• Decision Process: The decision to adopt an AI OCR solution stemmed from a comparative review of Nanonets, Laserfiche, and EasyOCR. The firm recognised the need for an AI consultant and opted for custom document processors using Google Cloud Platform and AWS Recognition technologies.
• Implementation Steps: The implementation began with isolating and testing OCRs on the most challenging documents and sections, striving for a satisfactory F1 score and passing the Golden Test set review. The solution involved a multi-processor approach with AI decision-making at times.
  
  Tech Stack: GCP Document AI, ReTool, Vercel, OpenAI, Process St. Typeform

Results and Impact

• Outcomes: AI OCR technology has revolutionised the mortgage submission process, reducing the time from weeks to just one hour. Instant feedback during the document collection phase significantly improved conversion rates and submission accuracy, making P. Stealth Mortgage Brokers stand out in one of the world's most regulated environments.

Differentiation and Value

• Unique Value Proposition: The firm's approach is uniquely tailored to the Australian market, employing the latest AI-OCR releases. Combining AI decision trees and traditional coding logic simplifies the document processing pipeline.
• Best Practices and Lessons Learned: A key takeaway from the project is simplifying the Custom Document Extractor (CDE) to its most challenging task to ensure accuracy. The team emphasised the need to monitor the F1 score while increasing complexity and adding new processors and categorisers as needed.