Abstract: With the rapid development of edge computing and intelligent Internet of Things, embedded AI technology has shown broad application prospects in intelligent manufacturing, automatic driving and other fields due to its low power consumption and high real-time characteristics. To promote the deep integration of artificial intelligence teaching and embedded technology, a handwritten digit recognition teaching experimental case based on Zynq platform was designed and developed. The system is based on the classic MNIST dataset and has constructed two models, Fully Connected Network (FCNN) and Convolutional Neural Network (CNN), respectively. High Level Synthesis (HLS) technology is used to achieve hardware acceleration. Through parallel computing optimization on the Programmable Logic (PL) side, the model inference delay was significantly reduced while maintaining model accuracy, ultimately achieving recognition accuracies of 98.20% and 99.05% in the Zynq system, respectively. Through the full process practice of “algorithm design hardware optimization system deployment”, teaching cases enable students to deeply master the key technologies of embedded artificial intelligence development, cultivate their software hardware collaborative design ability and engineering optimization thinking. The feedback from teaching practice shows that the experimental cases have effectively improved students' ability to solve complex engineering problems in the edge computing scene, and provided a replicable innovative paradigm for AI curriculum reform.