<p>Asynchronous Video Interviews (AVIs) are transforming recruitment, necessitating accurate recognition of top personality traits for effective candidate screening. While current models achieve high overall accuracy, they often fail to identify primary traits due to the central tendency of personality scores and the limitations of Mean Squared Error (MSE) optimization. This study introduces Light-Rank, a lightweight multimodal model specifically designed for accurate top-k trait recognition in resource-constrained environments. The architecture integrates text, audio, and visual features using pre-trained mobile backbones with frozen weights, training only the multimodal fusion regression head to ensure inference efficiency. Six modality combinations and three ranking loss functions (point-wise, list-wise, pair-wise) are evaluated on the ChaLearn First Impressions V2 dataset. Results demonstrate that the MSE + Cosine (pair-wise) loss significantly outperforms other configurations in identifying dominant traits. Mobile deployment evaluations confirm that the model achieves efficient real-time inference with minimal accuracy degradation after quantization. This approach provides a practical solution for streamlining recruitment workflows and enabling personalized candidate assessment in low-resource settings.</p>