A Runtime Anomaly Detection and Fixing Framework Using LLMPowered Separate guided Graph Attention Recurrent Compact Conviformer for Web Applications

Abstract

Modern web applications, particularly microservice-based, face numerous runtime anomalies from unexpected inputs, dynamic routing, permission conflicts, and misconfigured services. Existing detection systems also lack real-time adaptability and automated fixes, requiring manual intervention and reducing system resilience. To address these limitations, this research proposes a comprehensive runtime anomaly detection and fixing framework for Python web applications built on FastAPI. It employs a Separate guided Graph Attention Recurrent Compact Conviformer (SGARCC), which integrates graph-based modelling with recurrent attention mechanisms, enabling the detection of long-range temporal dependencies and relational patterns in log sequences. Hyperparameter optimization is achieved through the Bighorn Sheep Optimization Algorithm (BSOA), which dynamically tunes learning rates, dropout probabilities, and attention thresholds to enhance accuracy and reduce inference latency. Upon anomaly detection, Local Interpretable Model-agnostic Explanations (LIME) generate feature-level visualizations to improve transparency. large language models, specifically Gemma3, then provide context-aware, semantically correct code fixes, leveraging error logs, impacted code