LoanBoard AI - Intelligent Loan Boarding Engine

AI-powered tool that reduces loan boarding from 70 minutes to 5 minutes by comparing Term Sheets with Facility Agreements, instantly detecting discrepancies and generating servicing-ready data.

Comment

Inspiration

The infamous Revlon case was our wake-up call. In 2020, Citibank accidentally wired $900 million due to a confusing user interface and manual error during a routine loan operation. This wasn't a technology failure—it was a process failure.

We discovered that loan boarding—the process of entering data from signed PDF loan agreements into servicing systems—takes an average of 70 minutes per loan and is still largely manual. Even worse, critical terms often "drift" between the initial Term Sheet and the final Facility Agreement, creating compliance nightmares and financial risk.

We asked ourselves: What if AI could do this in 5 minutes with zero errors?

What it does

LoanBoard AI is an intelligent document comparison engine that:

  1. Ingests both the Term Sheet (initial agreed terms) and Facility Agreement (final legal document)
  2. Extracts key commercial terms using Gemini AI (borrower, facility amount, interest rates, covenants, definitions)
  3. Compares the documents side-by-side with visual highlighting
  4. Detects discrepancies instantly (e.g., "Term Sheet says 3.5% margin, Agreement says 3.25%")
  5. Generates servicing-ready JSON output for direct integration with loan management systems

How we built it

We built LoanBoard AI as a modern React + TypeScript application:

  • Frontend: React 19 with TypeScript, styled with Tailwind CSS and custom glassmorphism effects
  • AI Engine: Google Gemini 2.5 Flash for document analysis and data extraction
  • Architecture: Vite for fast development, modular component structure
  • Document Processing: PDF ingestion with Base64 encoding for AI analysis

The core innovation is our comparison algorithm that semantically matches fields across documents and calculates confidence scores for each extraction.

Challenges we ran into

  1. Document variability: Loan documents have no standard format. We solved this with flexible prompting and confidence scoring.
  2. Covenant complexity: Financial covenants like "Leverage Ratio < 4.0x" required custom parsing logic.
  3. Real-time comparison: Matching extracted fields across two documents with different structures was tricky.
  4. Demo without real documents: We built a comprehensive demo mode with realistic mock data.

Accomplishments that we're proud of

  • 10x speed improvement: From 70 minutes to under 5 minutes
  • Visual discrepancy detection: Color-coded side-by-side comparison that anyone can understand
  • Production-ready output: JSON export directly compatible with loan servicing systems
  • 98%+ AI confidence: Gemini 2.5 Flash delivers highly accurate extractions

What we learned

  • The loan market is ripe for digital transformation—many processes are still PDF and email-based
  • AI document understanding has matured significantly—what seemed impossible 2 years ago is now achievable
  • User experience matters as much as accuracy—banking professionals need clear, visual tools

What's next for LoanBoard AI

  • API Integration: Direct connections to major loan servicing platforms (Finastra Loan IQ, FIS)
  • Bulk Processing: Handle entire portfolios, not just individual loans
  • Audit Trail: Full compliance logging for regulatory requirements
  • CDM Export: Output in FINOS Common Domain Model format for industry standardization

Built With

Share this project:

Updates