International Journal of Scientific Research and Engineering Development

In production machine learning systems, data drift and concept drift pose severe challenges to model reliability, particularly in high-stakes domains such as credit card fraud detection. Standard static models suffer from performance degradation when transaction behaviors and attack vectors evolve over time. Traditional full-retraining strategies are computationally expensive and prone to catastrophic forgetting. In this paper, we propose a modular, production-grade MLOps continuous monitoring and auto-retraining framework powered by a novel Global-Local Ensemble (Hybrid Classifier) architecture. The system employs twosample Kolmogorov-Smirnov (KS) statistical testing across 29 domain features to detect statistical distribution shifts while excluding non-stationary temporal features to eliminate false-positive alerts. A 3-Tier Decision Policy (GREEN, YELLOW, RED) controls retraining triggers to suppress noise-driven compute overhead. Upon detecting significant drift or accuracy drops, a lightweight Local Expert (30% weight) is retrained adaptively over a 5,000-sample sliding window buffer, while a static Global Expert (70% weight) preserves long-term historical fraud patterns. Experimental evaluation on 284,807 transactions demonstrates that when facing a novel attack pattern, the static global model accuracy degrades to 98.60%, whereas the proposed Global-Local Ensemble restores accuracy to 99.42% (+0.82% recovery) in sub-second execution time.

Dr. Basavraj Neelgar, Malika Harmaien, Manasa L Hegde, Rakshitha M R, Samrat M, "AI-Powered Model Monitoring System with Data Drift Detection and Automated Retraining for Machine Learning Model" International Journal of Scientific Research and Engineering Development, V9(4): Page(1076-1083) July-August 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.