This paper presents an intelligent evaluation model for graduate employment quality based on cross-modal feature learning. Addressing the limitations of traditional assessment methods—including static weighting schemes, inefficient data collection, and underutilization of unstructured data—propose a novel multi-modal fusion framework that dynamically integrates structured employment metrics with unstructured textual feedback through a hybrid deep learning architecture. Our model employs a dual-path feature extraction network: a 5-layer fully connected network for structured data and a fine-tuned BERT-base-Chinese model for text, followed by a 4-head attention mechanism for cross-modal fusion. Key innovations include a discipline-specific adapter for dynamic weight allocation and an interpretability module that achieves 0.88 SHAP value consistency. Evaluated on a longitudinal dataset of 2843 graduates, the model demonstrates superior performance compared to baseline methods, while maintaining efficient inference. Practical applications show a 15 × acceleration in assessment speed and a 31.2% improvement in career guidance accuracy. This work bridges educational data mining and AI-driven evaluation, offering a scalable solution for higher education institutions to monitor and enhance graduate employability. Future directions include long-term career tracking and lightweight mobile deployment.