Abstract: To address the problems of low manual efficiency, inconsistent standard implementation, and difficult identification of hidden risks in railway contract auditing, this paper designed an intelligent review system for railway contracts based on the collaboration of large language models and multi-agent, and elaborated the system architecture, intelligent review procedures, and core functions. The paper adopted the Qwen2.5-VL-72B and Qwen3-32B models, combined with browser plug-in technology and retrieval-augmented generation technology, and realized the full-process processing covering unstructured data extraction, structured parsing of key elements, and automatic identification of compliance risks for contracts. Application results show that the proposed system reduces the average review time of a single contract from 30 minutes of manual work to 3~4 minutes. It also achieves a significant improvement in the accuracy of clause conflict detection and responsibility boundary definition, and effectively meets the standardization and intelligent development demands of contract management for railway enterprises.