Abstract: To address the longstanding challenges in engineering CFD—namely the high entry barrier, complex workflow configuration, heavy dependence on expert experience, and the high cost of convergence tuning—this work develops ZHILIU, a traceable, reusable, and fully localizable AI-CFD multi-agent platform for accelerated simulation. The platform adopts a multi-stage orchestration centered on a large language model and integrates retrieval-augmented generation (RAG) with an engineering knowledge graph to transform natural-language requirements into structured tasks. Within a coordinated multi-agent framework, it performs automated anomaly diagnosis and self-healing parameter tuning, while providing unified interfaces for geometry preparation, meshing, solvers, and post-processing. Local deployment, permission management, and audit logging further ensure security and compliance. The system enables end-to-end automation from requirement parsing and model setup to solving and visualization reporting. Mesh or convergence anomalies can be automatically detected and corrected through parameter optimization, while human intervention can be incorporated at any stage to support flexible human-AI collaboration and diverse simulation needs. The platform also supports concurrent execution and GPU acceleration to increase throughput and substantially enhance overall computational efficiency. Standard benchmark comparisons between the ZJILIU platform and conventional CFD workflows were conducted. The results show that ZHILIU significantly reduces manual intervention by 65.26% and cuts operational steps to just 1/8 of the conventional approach. It also demonstrates an advantage in solution stability. Overall, ZHILIU provides a practical and deployable pathway for transitioning CFD from tool-centric workflows to intelligent, collaborative simulation. This study is primarily based on the domestic open-source CFD platform (NNW-PHengLEI), and invokes its core solver module. It provides a feasible path for transforming CFD from tool-driven to intelligent collaborative approaches. The platform improves efficiency without sacrificing accuracy, and is suitable for both research and industrial applications.