Hybrid LLM + ML Document Classification with Error Recovery Designed a hybrid AI system for classifying domain-specific documents under noisy and highly imbalanced conditions. Combined XGBoost, transformer models (BERT-family), and retrieval-augmented LLM reasoning to improve accuracy and robustness. Introduced an “anti-breakage” mechanism to prevent LLM corrections from degrading correct predictions. Built end-to-end pipelines including preprocessing, feature engineering, model training, evaluation, and LLM fallback logic. Optimized for real-world reliability and edge-case handling. Impact: ~89% accuracy, improved robustness, reduced manual effort.