Purely data-driven AI has achieved remarkable success in engineering applications, yet it remains fundamentally limited by data scarcity, poor extrapolation, and a lack of physical interpretability. In this talk, we present a knowledge-guided AI framework that integrates physics knowledge and human expertise into machine learning for robust and data-efficient engineering design. We first introduce Physics-Informed Machine Learning (PIML), highlighting physics- informed neural networks (PINNs) that embed governing equations and physical constraints to enable reliable inference and inverse design from extremely limited data. We then extend this approach to human-knowledge-guided AI using multi-agent large language models (LLMs), which orchestrate simulations, optimization tools, and expert knowledge for user-friendly and autonomous manufacturing workflows. Case studies in structural optimization and injection molding demonstrate the effectiveness of this paradigm for next-generation intelligent manufacturing.