AI Document Processing for Digital Lending
How a Mumbai-based NBFC reduced loan underwriting from 8 days to under 24 hours with an AI-powered document verification system.
Karma FinCorp, a Mumbai-based NBFC serving SME borrowers across Maharashtra and Gujarat, was processing loan applications through a manual 14-document verification workflow that took 8-10 days. VictorLabs built an AI document processing pipeline that automated verification, fraud detection, and compliance checks, reducing turnaround to under 24 hours while meeting RBI audit requirements.
Underwriting time
8 days → 24 hrs
Rejection rate
22% → 8%
Throughput increase
3x
Languages supported
4
Context
Karma FinCorp is a mid-sized NBFC focused on SME lending in western India, processing 2,500+ loan applications monthly. Their underwriting team of 18 manually verified 14 document types per application — PAN cards, GST returns, bank statements, ITRs, business licenses, and property documents. The RBI mandates strict audit trails for all lending decisions.
Problem
Each loan application took 8-10 days to underwrite. 14 document types meant 14 different verification workflows. Documents arrived as scanned PDFs, phone-camera photos, and email attachments. GST returns were in Hindi, bank statements in English. 22% of applications were rejected due to incomplete documentation — many of which could have been caught earlier. The RBI had issued observations about inconsistent documentation in their last audit.
Constraints
All data processing had to remain within India to meet RBI data localization requirements. Document formats varied wildly — from crisp scanned PDFs to blurry smartphone photos. Forms and certificates in Hindi, Marathi, and Gujarati needed to be understood. The audit system required explainability for every automated decision. Integration with the existing loan management system (built on a legacy Java stack) was mandatory.
Approach
We started with a discovery sprint at their Mumbai office, shadowing the underwriting team to understand the real workflow — not just the documented process. We identified 3 highest-volume document types (bank statements, GST returns, PAN cards) for Phase 1. Built a classification + extraction pipeline that handled multilingual documents. Designed an explainability layer that annotated every extraction with confidence scores and source references for audit compliance.
Solution
The system classifies incoming documents, extracts structured data using vision models and multilingual LLMs, cross-validates fields against external databases (PAN verification, GSTIN lookup), and flags discrepancies for human review. A dedicated review interface shows the extracted data side-by-side with the original document, highlighting low-confidence fields. Verified applications are posted to the loan management system with a complete audit trail. The system handles Hindi, Marathi, Gujarati, and English documents natively.
Architecture
Deployed on Indian cloud infrastructure (Mumbai region) for data localization compliance. Backend: Python FastAPI services for document processing pipeline. Document intelligence: Azure Document Intelligence for OCR, fine-tuned Indic-language models for extraction. PostgreSQL for audit logging and transaction data. Frontend: React-based underwriter workspace with review interface and dashboard. Integration layer connects to the legacy loan management system via a REST adapter.
Results
Underwriting turnaround dropped from 8-10 days to under 24 hours. Application rejection rate fell from 22% to 8% — the system caught incomplete applications at submission rather than at final review. The underwriting team now processes 3x volume with the same headcount. RBI audit observations were resolved in the next review — the audit trail and decision explainability features were specifically commended. Karma FinCorp expanded to Gujarat and Rajasthan in the following quarter.
Lessons
Shadowing the team in person was invaluable — the real workflow had undocumented steps and informal checks that the documented process missed. Building for multilingual support from day one was essential — half the documents were in Indic languages. The explainability feature (showing exactly why a field was flagged or extracted a certain way) became the most valuable feature for both the underwriting team and the compliance department.
“Our loan underwriting took 8-10 days with manual verification across 14 document types. VictorLabs built an AI document pipeline that reduced it to under 24 hours. We disbursed 3x more loans this quarter with the same team. The RBI compliance team signed off on the audit trail in the first review.”
— Vikram Malhotra, CTO, Karma FinCorp