Enhancing Software Development Efficiency and Code Quality Using AI-Driven Transformer-Based Code Generation

Abstract

In modern software development, efficiency and code quality are critical for project success. Traditional coding approaches often struggle with scalability, maintainability, and error reduction. This paper explores the use of AI-driven transformer-based code generation, specifically leveraging the hybrid integration of CodeT5 and TreeLSTM models. CodeT5, a transformer-based model, facilitates contextual code understanding and generation, while TreeLSTM processes hierarchical code structures like Abstract Syntax Trees (ASTs) to enhance syntactic correctness and structural integrity. A Kaggle dataset comprising multi-language code snippets is used for training, with Byte Pair Encoding (BPE) applied for tokenization. Additionally, model pruning techniques optimize performance by reducing computational overhead without sacrificing accuracy. Comparative analysis of sorting and search algorithms, including Bubble Sort, Quick Sort, and Binary Search, highlights the importance of algorithm selection in execution efficiency. Experimental results demonstrate that the hybrid CodeT5 + TreeLSTM model significantly improves code generation, refactoring, and optimization by reducing redundant computations and improving execution time. The proposed approach not only enhances coding efficiency but also ensures improved software maintainability and scalability. However, challenges remain in terms of model interpretability, dataset biases, and real-world adaptability. Future research will focus on refining AI-based code generation for broader applications, improving generalization across different programming paradigms. This study contributes to advancing automated software development by leveraging AI techniques, thereby addressing the growing complexities and demands of modern programming environments.