CodeQual-Agent: An Intelligent LLM-Agent Framework for Automated Software Quality Assessment with Explainable Predictions and Real-Time Analysis

Software quality assurance requires scalable solutions combining the deterministic precision of static analysis with the contextual reasoning of semantic understanding. This paper presents CodeQual-Agent, a hybrid neural-symbolic framework bridging traditional metrics-based assessment with Large Language Model (LLM) capabilities through an intelligent orchestration architecture. The framework addresses three fundamental challenges: (1) integration of symbolic features with neural reasoning, (2) efficient distillation of quality-assessment expertise from high-capacity teacher models into deployable student agents, and (3) explainable predictions that build developer trust. CodeQual-Agent employs a Virtual SonarQube Instance as its symbolic foundation, extracting objective metrics (cyclomatic complexity, code duplication, maintainability indices) that serve as attention-weighted anchors grounding LLM reasoning. Through knowledge distillation from GPT-4 and Gemini Pro teachers to a Gemini 2.5 Flash student, the framework achieves 94.0% precision at 450ms inference latency, maintaining 96.8% of teacher accuracy while delivering a 2.6× speedup. A Human-in-the-Loop (HITL) active learning pipeline reduces false positives by 17% over three update cycles. SHAP-based explainability achieves 87% alignment with developer intuition. Evaluated on 150,000 samples across three complementary datasets, real-world deployment demonstrates a 14% reduction in production defect escape rates and 83% developer adoption.