Large Model Private Deployment and Risk Control Application at a Joint-Stock Bank

Project Background

A joint-stock bank faced significant challenges in credit approval and compliance review. Traditional credit approval relied on manual review of borrower information, credit reports, and financial data, taking 3 days per application with a risk miss rate of 3%. Meanwhile, the volume of regulatory compliance documents was massive and frequently updated, making it difficult for the compliance team to efficiently analyze documents and identify risks. The bank had strict data security requirements—all business data must not leave the premises, and AI inference must be performed locally.

Core Pain Points

Solution

Private Large Model Deployment

Deploy the Qwen2.5-72B large model on the bank's local GPU cluster (8×A100), using the vLLM inference framework to optimize throughput. All model inference and data processing occur within the bank's internal network, ensuring zero data leakage and full compliance with CBIRC data security regulations.

Intelligent Credit Risk Review

Build a credit risk review assistant based on the large model that automatically parses borrower information, credit reports, and financial statements, cross-verifies information consistency, identifies potential risk points, and generates review reports. Approval time was reduced from 3 days to 4 hours, and the risk miss rate dropped from 3% to 0.5%.

Intelligent Compliance Document Analysis

Develop an intelligent compliance document analysis system that supports automatic interpretation of regulatory documents, internal policy compliance checks, and impact assessment of policy changes, freeing the compliance team from repetitive reading and analysis tasks.

Results

Tech Stack

Qwen2.5-72B, vLLM inference framework, NVIDIA A100 GPU cluster, LangChain, Python, Kubernetes, isolated deployment on bank intranet