08/06/2026
✨Another week we are starting with the doctoral colloquium✨
Munura Maihankali with the topic From Code Decay to Verified Transformation: A Semantic Correctness Assurance Framework for Agentic AI-Driven Refactoring of Large-Scale Software Systems
Abstract: Software systems operating at scale e.g. enterprise applications, distributed platforms, and long-lived production codebases inevitably accumulate structural decay over time. In large-scale systems where thousands of interacting classes, services, and modules evolve under continuous delivery pressure, code gradually develops decay symptoms such as inflated methods, densely coupled classes, violated architectural boundaries; all of which erode maintainability and long-term evolvability. Refactoring, the discipline of restructuring existing code without altering observable behaviour, is the established first response mechanism by which software recover from structural degradation. This study is targeted at refactoring large-scale software systems hence a robust approach requires reliable metrics which are the only objective mechanism for monitoring code health, diagnosing structural decay, and verifying whether a refactoring has genuinely improved the system. This research employs a five-family metric suite spanning cohesion, coupling, complexity, readability, and architectural health, measured systematically both before and after every refactoring transformation to provide empirically grounded semantic correctness verification. Large Language Models and agentic AI systems now offer refactoring automation at scale yet automation without verification is dangerous. LLMs are probabilistic generators which are prone to hallucination, context-window blindness, and silent semantic violations. This research proposes a multi-layer verification architecture: smell-aware ordering of refactoring operations, the five-family metric suite, Graph-RAG-enhanced contextual awareness (LlamaIndex, ChromaDB, Neo4j), LangGraph orchestration, and a rejection-policy filter which has yielded tentative pilot findings and novel contributions to the refactoring domain.
More about all colloquia here 👉https://dai.fmph.uniba.sk/w/Doctoral_Colloquia/en