Geração automática de sistemas backend com o suporte de IA generativa seguindo a arquitetura limpa

Abstract

In this work, we investigated the potential contribution of automatic code synthesis technologies, particularly OpenAI’s GPT-4, to the maintenance and adherence to best practices in software architecture in startups. Given the characteristic of these companies to operate in environments of rapid change and innovation, but with limited resources, practices such as unit testing and documentation are often neglected. Conversely, we emphasize the importance of such practices for their contribution to the maintainability and scalability of applications. As a means to reconcile the fast pace of development with the need for good practices, we proposed the use of generative language models (GLM), specifically GPT-4, for code generation following the principles of clean architecture, a set of concepts defined by Robert C. Martin for developing scalable and maintainable projects. The methodological approach was a combination of qualitative and quantitative analysis, focused on the exploration and adaptation of prompts for code generation and the development of practical exemplifications in various programming languages. Notably, three projects in C#, JavaScript, and Python were produced, which were evaluated according to metrics of abstraction, instability, and adherence to the Main Sequence - key concepts in maintaining clean architecture. The results indicated that, despite the potential of the proposed technology to accelerate development and promote adherence to good practices through automation, there are significant gaps in GPT-4 ability to generate code fully aligned with clean architecture and executable without manual intervention. Problems related to inconsistency in the project structure and the integrity of the generated code were observed, suggesting that, while the tool offers a promising foundation for enhancing efficiency in less complex projects, its applicability in complex and diverse contexts still presents challenges. Therefore, it is concluded that the use of GLMs like GPT-4 in the automatic generation of code represents a valuable auxiliary tool for startups in software development. However, the need for manual adjustments in the code and the assurance of full adherence to recommended software architecture practices reinforce the idea that such technologies should be seen as complementary to human work and not as complete substitutes. For future work, it is recommended to deepen the investigation of GLMs specialized in code generation and to expand the experiments to encompass a wider range of programming languages and frameworks, aiming to maximize the applicability and effectiveness of this innovative approach.