Production-minded systems across LLM applications, AI for software engineering, full-stack development, and machine learning.
A production-ready knowledge system with a six-stage RAG pipeline, parallel retrieval, TTL model caching, SSE streaming, and containerized deployment. Retrieval recall improved by about 40% and first-token latency fell by about 60%.
An open-source repository health and lifespan prediction system built on more than 100,000 GitHub repositories. A hybrid deep-learning model reached R² 0.95+ and won a national third prize in the CCF Open Source Innovation Competition.
An editable multimodal poster generator that coordinates Planner, Visual, Layout, and Critic agents with a dynamic object-oriented layout engine and a RAG knowledge base.
A comparative evaluation framework for American Sign Language alphabet recognition across CNN, Liquid Neural Network, and Vision Transformer architectures, including robustness and interpretability analysis.