Experiments on Code Generation for Web Development using LLMs: A Comparative Study

This thesis explores how large language models (LLMs) can support web development, focusing on a comparison between DeepMind's Gemini models and OpenAI’s GPT models in generating code for both front-end and back-end tasks like HTML, CSS, JavaScript, and PHP. Organized into three main sections, the study covers how LLMs are being used in web development workflows, the foundations of these models, and a detailed comparison of GPT and Gemini in terms of code generation and evaluation performance. Using a series of experiments, this research evaluates the accuracy and accessibility of both models in handling various coding tasks. It also examines ways to improve model performance through prompt optimization techniques to achieve the most optimal output on the first attempt. An additional section focuses on an accessibility review, assessing Gemini's ability to detect accessibility errors and evaluate its effectiveness in code correction. The findings reveal differences between GPT and Gemini, showing how each model approaches code generation and error correction, as well as their support for web accessibility standards. By optimizing the prompts given to the models, this study demonstrates that LLMs can significantly elevate code quality and enhance web accessibility.