<p>Optical Character Recognition (OCR) has progressed widely for Indian scripts such as Bengali, Devanagari, and Urdu, but Kashmiri script remains underexplored. Handwritten cursive recognition in scripts like Nastaliq, used in Urdu and Kashmiri, is particularly challenging due to cursive connections, context-dependent character shapes, and handwriting variability. Existing methods often fail to capture crucial features like loops, dots (nuqta), and stroke details that distinguish visually similar characters. This study proposes a deep learning approach based on the Swin Transformer architecture, which employs windowed attention mechanisms and multi-scale processing to retain both fine details and holistic structures. The model addresses difficulties such as overlapping characters and diverse handwriting styles, adapting well to the intricacies of handwritten Urdu and Kashmiri text. A multi-scale windowing technique ensures that both fine details and larger character structures are captured, enabling the model to handle overlapping characters and varying handwriting styles. Experimental evaluation on four benchmark datasets (UNHD, NUST-UHWR, UHLD, and the newly introduced HKTD for Kashmiri) demonstrates that the proposed Swin Transformer approach achieves an overall recognition accuracy of 95.4% with a ligature error rate as low as 1.5–2.9% for Urdu datasets and 5.6% for Kashmiri text. The model requires less than 3 GB GPU memory and completes training in under 3&#xa0;hours, significantly outperforming existing CNN and hybrid Transformer-based baselines in both accuracy and computational efficiency.</p>