Optical Character Recognition (OCR) systems for Hindi script continue to produce significant word-level errors, despite high character-level accuracy largely due to visually similar glyphs and ligature complexity which makes segmentation of characters challenging. Post-processing can address this, but most existing correction modules are either context-free or black-box generative models, limiting precision and control. In this paper, we propose a hybrid model that combines confusion-aware candidate generation with a RoBERTa-based reranker for post-OCR word correction. Given a word-level OCR output from a CRNN model, we generate the top-10 candidate corrections using a Levenshtein search enhanced with a handcrafted confusion map. A binary RoBERTa classifier is then fine-tuned to re-rank these candidates using over 100k confusion-simulated word pairs. We evaluated the model on a set of 2,699 real OCR errors, achieving 59.54% Top-1 Re-ranking accuracy and correcting the 1,607 errors, thereby improving overall OCR accuracy from 97.70% to 99.07%. To our knowledge, this is the first system to apply confusion-aware reranking to Hindi OCR, demonstrating that lightweight language models can significantly improve OCR reliability for low-resource scripts.