A complex neurodevelopmental learning disability, dyslexia is typified by ongoing difficulties with visual phonological mapping, orthographic processing, and reading. The fine-grained analysis of handwritten letter formation patterns can be a reliable indicator of dyslexic traits, according to recent developments in deep learning-based computer vision. However, most existing research focuses on coarse screening (dyslexic versus non-dyslexic) and lacks methods for continuously and interpretably measuring the severity of dyslexia. In order to quantify the severity of dyslexia through sequence modeling, this paper presents a hybrid deep learning pipeline that simultaneously uses a ResNet50-based embedding extractor and You Only Look Once (YOLOv5) for granular letter-level detection. Whole handwriting pages worth of extracted letter embeddings and spatial-structural patterns are used to train a custom sequence regression model. Dyslexia handwriting images with three representative error categories-Normal, Reversal, and Corrected letters-are used to test the suggested framework. According to the results, the suggested severity regression framework achieves an average RMSE of 0.0146 and Pearson correlations above 0.967 across folds, while YOLOv5 achieves a nearly perfect mAP50 (0.995) across all classes. These results indicate that, within this synthetic setting, severity scoring can be modeled as a continuous, geometry-aware regression task rather than a purely binary classification problem. The synthetic pages are made out of character samples from several sources including public handwriting data sets – making sure that there is a realistic variability at the letter level with programmatic control over layout and annotations. Severity scores in this study are computed as the sum of the reversal and corrected-letter rates per page. All experiments use programmatically generated pages derived from real character samples.