This study introduces a novel fashion recommendation system that integrates visual and textual information using advanced deep learning techniques. The system analyzed a dataset of approximately 44,000 fashion items, each having image and text description. For textual data, the system employs bidirectional encoder representations from the transformer (BERT), while visual features are extracted using ResNet-50 model. Comparative analysis revealed that ResNet-50 achieved higher validation accuracy of 0.9895, outperforming other models, demonstrating its superior performance in visual feature extraction. These features were combined through a late fusion approach enhancing computational efficiency while maintaining accuracy. Unlike traditional systems that depend on user reviews or sales data, our approach operates independently of such inputs, making it highly applicable in warehouse-based applications where such data is unavailable during initial deployment. Recommendations are generated by constructing a similarity matrix to identify the top five products most similar to a given query item, reflecting attributes such as design, color, and material. The evaluation showed the system’s high accuracy in providing relevant recommendations with cosine similarity scores exceeding 0.99 for the top recommended products. However, limitations include reliance on pretrained models. Future research should explore more advanced architectures, real-time systems, and additional data modalities to improve the performance across various domains.