International Journal of Science, Engineering and TechnologyAuthors: R. Hanush Singh, Dr. G. Maragatham
Abstract: Software code review is a labor-intensive process that directly influences the reliability, security, and maintainability of modern software systems. Conventional automated review tools, including linters, static analyzers, and monolithic large language model (LLM) assistants, suffer from limited contextual reasoning, narrow analytical scope, and a lack of explainability. This paper presents CodeGenie, a multi-agent generative artificial intelligence (AI) framework that decomposes the code review task into five specialized analytical roles: syntactic correctness, security, performance, stylistic consistency, and documentation quality. Each agent is implemented as a prompt-conditioned instance of a foundation generative model and operates in parallel on a shared code artifact. A weighted decision-fusion mechanism aggregates per-agent verdicts into a unified quality index, while an adaptive feedback loop personalizes review granularity based on developer history. The framework is evaluated on a curated benchmark of 480 source files spanning Python, JavaScript, and Java, drawn from open-source repositories and synthetic defect injections. Experimental results demonstrate that CodeGenie achieves a defect detection F1-score of 0.892, a 17.3% improvement over a single-model baseline, while reducing reviewer cognitive load by 41% in a controlled user study with 24 developers. The system contributes to United Nations Sustainable Development Goals 4, 8, and 9 by democratizing access to high-quality code review, improving developer productivity, and strengthening software infrastructure. We discuss the system architecture, mathematical formulation, empirical results, threats to validity, and future research directions toward self-improving agentic software engineering assistants.
