Abstract: High safety risks exist at construction sites, making pre-shift safety education a critical measure for accident prevention. However, traditional education methods relying on manual roll calls and paper records suffer from low efficiency, error-proneness, and difficulties in supervision. To address these issues, this paper designs and implements an intelligent pre-shift education supervision system based on AI facial recognition technology for crowds. The system integrates a camera module, data acquisition module, voice recognition module, record generation module, and early warning module. It employs a crowd aggregation triggering algorithm and convolutional neural networks to achieve automatic identity verification, while collecting voice and environmental data in real time to enable structured storage of educational content and automatic report generation. The system was further applied in the reconstruction project of Yaotai Village, Yuexiu District, Guangzhou. The application demonstrates that under normal lighting conditions, the system achieves a facial recognition accuracy rate of 100% with an average response time of 1.8 seconds and a Data Quality Index (DQI) of 0.92. Under low-light and rainy conditions, the recognition rates are 91% and 68%, respectively, while the DQI remains above 0.75, and the Comprehensive Warning Index (CWI) does not exceed the threshold of 0.3, indicating good system robustness. In a practical scenario test involving 50 workers, the system effectively prevented attendance fraud, improved the standardization of education records by approximately 40%, and is estimated to save about 150,000 yuan in annual labor management costs. This study demonstrates that the system significantly enhances the automation and standardization level of pre-shift education, providing a feasible technical solution for safety education management at construction sites.