Development and Evaluation of an LLM-Based Tool for Automatically Building Web Applications

Abstract

In this thesis, we present Kodless, a platform that enables users to automatically build web applications from natural language descriptions without requiring them to write, review, or debug the generated code. Kodless structures applications using concept design, a theory which views software as a collection of interacting yet independent units of functionality mapping to human behavior patterns. The platform leverages large language models to generate functional backend code, combining concept design principles with a robust framework for developing concept implementations and integrating them into a standardized application architecture. To evaluate Kodless's performance, we conduct a study in which we use the platform to develop an application through an iterative prompt refinement process. We argue that the case study illustrates the importance of concept-driven prompt engineering and offer guiding principles for designing effective prompts. Furthermore, this thesis contributes improvements to the Kodless platform, including extended support for MongoDB integration and the automatic generation of a frontend testing client. We also introduce a frontend code generation assistant to enable automatic generation of reactive user interfaces. Ultimately, Kodless represents a promising path towards changing how we approach AI driven software design and development.