Traditional psychological assessment methods, such as questionnaires and interviews, are prone to subjective masking, making it difficult to accurately capture individuals’ true psychological states. As a projective tool, electronic sandboxes allow users to reveal underlying psychological characteristics through object selection and spatial arrangement. However, current sandbox analysis relies heavily on manual interpretation, which lacks automation and objectivity. This study proposes an intelligent psychological assessment system that integrates Large Language Models (LLMs) and computer vision techniques to automatically identify sandbox features and annotate psychological keywords. The system consists of a data acquisition module, a feature extraction module, and a psychological keyword annotation module. Furthermore, an experimental framework is designed to validate the system’s accuracy and consistency by integrating behavioral, physiological, and subjective data. The proposed system provides an innovative approach to automated psychological analysis, overcoming traditional assessment limitations. Future research will focus on optimizing algorithms, expanding datasets, and validating its applicability across various psychological scenarios.