Automating multiple choice grading from unconstrained smartphone images is challenging due to shadows, perspective distortion, and template dependency, while machine learning approaches suffer from class imbalance and cell level data leakage. This paper proposes a robust, template independent hybrid OMR pipeline to address these vulnerabilities. The framework implements a dual path cropping engine using YOLOv5 coordinates backed by an intelligent grid fallback. To eliminate shadows, localized truncated thresholding (THRESH_TRUNC) and min max normalization are applied. For each cropped cell, we engineer a five dimensional morphological and edge feature vector comprising grayscale mean, minimum, standard deviation, ink fill ratio, and Sobel gradient edge density where Sobel edge density uniquely distinguishes hand drawn cross outs from filled marks. Crucially, a strict stratified document level splitting protocol isolates entire student sheets to eliminate feature leakage. Trained via H2O AutoML, the champion Gradient Boosting Machine (GBM) model achieved a validation log loss of 4.57 ×10^(-8). Evaluated on an independent, strictly isolated test set of 1,575 cells (1,556 empty, 8 filled, and 11 crossed out), the model achieved 100.00% accuracy, a 1.00 macro average F1 score, and a perfect ROC AUC, alongside a 0.00% False Negative and False Positive Rate, establishing a secure, fair, and deployment ready educational grading system................... Keywords:.................. Optical Mark Recognition, YOLO, H2O AutoML, Truncated Thresholding, Sobel Edge Density, Stratified Split, Student Fairness.