On-premise LLMs
Arabic and English open-source models deployed inside the bank — applicant documents are never transmitted to an external service.
Banking · Qatar Flagship engagement
Agentic loan officer
A multi-agent workflow that reads salary certificates, cross-checks pay slips, and computes eligibility for a leading retail bank — on open-source LLMs that never leave the bank's walls.
14→4 min verification time per application
48 h faster time-to-approval on average
97% precision flagging tampered documents
5 FTE review capacity returned to complex cases
The challenge
What was in the way.
- Loan officers spent 10–20 minutes per file manually reading salary certificates, pay slips, and wage-protection statements to confirm income and benefits.
- Retail demand reached up to 3,000 applications a day — manual checks slowed time-to-yes, frustrated customers, and inflated back-office headcount.
- Inconsistent document formats and occasionally edited PDFs raised compliance risk under central-bank guidelines.
What we built
The system, in brief.
Weekly learning loop
Officers correct any extraction; the edits retrain the models on-premise every week.
LOS integration
One-page summaries with confidence scores pushed straight into the loan origination system and an underwriting dashboard.
Fraud detection
Certificate totals compared against pay-slip history and wage records — mismatches and tampering flagged before underwriting.
The agents
Who does what.
- 01
Document-ingestion agent
Receives PDFs and images from the loan portal and branch scanners.
- 02
Vision-OCR agent
High-accuracy extraction of text
- 03
Semantic-parser agent
Maps fields — basic salary
- 04
Cross-check agent
Compares certificate totals with the last three pay slips and wage-protection records; flags mismatches or tampering.
- 05
Eligibility calculator
Computes debt-burden ratio
- 06
Report-writer agent
Produces a one-page summary with confidence scores for the underwriter.
Outcomes
What changed.
- Verification time per application fell 70% — from 14 minutes to 4.
- Average time-to-approval improved by 48 hours, lifting customer NPS by 12 points.
- Five FTEs of manual review capacity re-allocated to complex credit cases.
- Fraudulent or altered salary documents detected with 97% precision in acceptance testing.
Client referenced by sector and country · detailed references on request