Treffer: Data science in Java/JavaScript programming environment for AI-assisted programming

With the introduction of function calling in Large Language Models (LLMs), the capabilities of the models can be customized and extended beyond text generation to the invocation of custom functions created by developers. This project investigates the effectiveness of integrating custom LLM Functions that are equipped with code compilation and execution capabilities via Judge0, to improve the code generation process. The implementation is done entirely through the Nemobot web application, facilitating a smoother and more manageable development process. Various LLMs such as GPT-5 mini and DeepSeek R1 0528, including open-source LLMs such as Qwen3 Coder and Kimi K2, were used for the evaluation. The code correctness and performance of the LLMs were assessed using the datasets from HumanEval and Mostly Basic Python Problems (MBPP) to examine whether the results are consistent. The findings provide insights into the advantages and limitations of the integration, as well as the importance of choosing a LLM that possess strong programming capabilities. The developed system serves as a great foundation for further research and development on the use cases of function calling, and the possibilities of developing advanced Application Programming Interfaces (APIs) to integrate with LLMs to enhance their code generation capabilities. ; Bachelor's degree