AI-Powered Document Vetting for Credit Risk Using Open Source LLMs

Quick Summary

Challenge

A financial institution needed to streamline the manual and error-prone process of validating bank statements for credit checks, including transaction history and account eligibility.

Solution

Tatras Data developed a GenAI-powered system using open-source LLMs and a RAG-based extraction pipeline to automate data validation, with enhanced privacy, cross-format compatibility, and cost efficiency.

Result

Turnaround time dropped with no compromise on accuracy.

Tech Stack

AI: Open-source LLMs (LLaMA, others) prompt engineering for validation | ML: Transaction classification Entity extraction Embeddings | Data & Retrieval: RAG system Regex for date validation | Dev: PyTorch TensorFlow FastAPI LangChain LlamaIndex | Security: Fully on-premise Private data pipeline with audit and access controls

The Challenge

Validating customer bank statements for credit eligibility is a time-consuming task involving multiple checks:

Minimum balance requirements
Monthly transaction patterns
Recent account activity over 3+ months

For this financial institution, the process was not only labor-intensive but also inconsistent and costly.

They needed an automated solution that could interpret structured and semi-structured financial documents while respecting strict data privacy mandates.

A Day in the Life: Before Our Solution

An underwriter reviews a scanned bank statement, checking for minimum balance compliance, recurring transactions, and gaps in account activity.

They manually copy data into spreadsheets, use filters to review dates and values, and make eligibility decisions based on custom logic — repeated across hundreds of applications per week.

The process is slow, vulnerable to oversight, and impossible to scale efficiently.

Pain Points:

Manual vetting was slow, costly, and inconsistent
High error rates due to manual date and value checks
Varying bank statement formats complicated parsing
Compliance rules required all data to remain on-premise
No automated cross-checking for transaction continuity

Solution

1. Core Innovation

Tatras engineered a GenAI-powered document vetting system using open-source LLMs and a retrieval-augmented generation pipeline to extract, validate, and verify key eligibility signals.

Core Components:

RAG-Based Data Extraction
Combines document parsing with embeddings to extract transactions, balances, and account metadata.
Validation Engine
Checks extracted values against credit rules (e.g., 3-month average balance, transaction frequency).
Regex-Based Date Normalization
Handles diverse date formats for transaction history validation with high precision.
Privacy-First Deployment
All processing remains within the client’s infrastructure, ensuring regulatory compliance.

2. Key Features

RAG + LLM Extraction: Accurate parsing of bank statements
Automated Eligibility Checks: Based on credit and balance criteria
Cross-Format Compatibility: Handles multiple statement layouts
Data Privacy at Core: On-premise, audit-friendly architecture
FastAPI Deployment: Lightweight, REST-ready system integration

3. Workflow Integration

The system is deployed as an internal API layer that processes uploaded statements in real time.

Once a bank document is uploaded, the system extracts relevant data, runs validation checks, and returns a decision flag along with traceable results — reducing manual review to exceptions only.

Outcomes

✅ Enhanced vetting system speed and efficiency

🔍 Robust handling of varied transaction formats and layouts

📉 Reduced manual effort and associated risk

💸 Significant cost savings using open-source LLMs

🔒 Fully private deployment that meets internal compliance and security standards