Developed a high-efficiency intelligence pipeline system for B-end enterprises, realizing a fully automated commercial intelligence closed loop. Led the development of the service layer with FastAPI and SQLAlchemy, ensuring high concurrency and robustness. Utilized Celery for distributed task queues. Re-architected multi-agent collaborative workflows, using Cluster Agent with local lightweight BGE-M3 for text representation and Scikit-learn DBSCAN for unsupervised clustering. Developed a data panel with React 19 and Vite 8 for B2B systems.

Utilized large models to generate digital human clusters with complex game-theoretic behavior to test and optimize subsidy strategies for takeaway platforms. Analyzed consumer behavior using multi-dimensional features and employed inverse probability weighting to mitigate observational bias. Built a heterogeneous multi-agent sandbox using conditional variational autoencoders and swarm intelligence. Developed a conservative reinforcement learning mechanism to prevent offline model abuse. Proposed a 'differentiated subsidy strategy' based on high switching costs and provided recommendations for further improvement.