FIN-SMART: A Unified AI Framework for Fraud Detection, Compliance Automation, and Customer Intelligence in Banking

This paper introduces FIN-SMART, a modular AI framework addressing three key challenges in fintech: real-time fraud detection, regulatory compliance automation, and intelligent customer support. The primary contribution is a confidence-aware routing mechanism in the CRM module that dynamically selects between lightweight NLP models and a generative large language model based on prediction confidence, achieving a 73% reduction in response latency while improving accuracy to 89%. The fraud detection subsystem integrates Isolation Forest with XGBoost, achieving 91.2% precision and 88.5% recall on transaction anomalies. The compliance pipeline processes unstructured financial documents using OCR and vector-space regulatory mapping, attaining 94% accuracy against RBI and SEBI guidelines. Experiments are conducted on a proprietary banking dataset (500K transactions) and a synthetically generated NLP corpus, with evaluation performed using 5-fold crossvalidation and 95% confidence intervals. Results demonstrate that the modular architecture enables real-time processing (sub-200ms latency) while maintaining robust performance across multiple financial intelligence tasks.