Adaptive anomaly detection in production systems: A versioned autoencoder framework

This paper presents a comprehensive, production-grade anomaly detection framework integrating PyTorch-based autoencoders, threshold tuning, preprocessing persistence, and real-time inference via FastAPI with Prometheus monitoring. The training pipeline employs stratified train/validation/test splits, standardization, early stopping, checkpointing, and TensorBoard telemetry. Evaluation comprises precision, recall, F1-score, AUC-ROC, confusion matrices, and score distribution analysis with visualizations. The deployment accommodates versioned model artifacts, per-version thresholds, and preprocessing reuse, minimizing operational complexity and ensuring reproducible performance. We demonstrate end-to-end automation for figure generation, document assembly, and artifact management, emphasizing maintainability and academic rigor.