LLM-driven KYC onboarding
From 3–5 days to under 24 hours, with customer data never leaving the bank.
Senior Data Scientist
Customer onboarding meant reading documents by hand, checking KYC details and assessing credit risk, which took 3–5 days per case. Any automation had to keep sensitive customer data inside the bank’s own infrastructure.
How it works.
Intelligent document extraction
Tesseract OCR with OpenCV preprocessing pulls fields out of scanned customer documents.
LLM field validation and case summaries
A language model checks extracted fields for consistency and writes a GenAI summary of each case for the reviewer.
Explainable credit-risk scoring
An ML credit-risk model (0.87 AUC) scores each applicant, with SHAP explanations behind every score.
Privacy and control by design
PII masking, human-in-the-loop review, and deployment on on-premise Linux servers through Jenkins CI/CD pipelines.
Turnaround fell from 3–5 days to under 24 hours and manual review dropped by about 60%, with auditable, consistent compliance workflows and sensitive data kept in-house.
My role
- Designed and built the onboarding pipeline in Python
- Built the credit-risk model and its SHAP explanations
- Set up PII masking, human review and on-premise CI/CD deployment
Stack
- Tesseract OCR
- OpenCV
- LLMs
- SHAP
- Python
- Jenkins
- Linux