Explainable AI Code Generation Assistant

Abstract

Large language models can generate and modify code for non-trivial program-ming tasks, but their explanations often remain generic or weakly connected to the implementation. When an agent edits several files, the developer must understand what changed and where each explanation is grounded in the code. The main objec-tive of this dissertation is to design, implement, and evaluate an explainable AI code generation assistant that connects natural-language explanations to source-code lines. The system was implemented as an IntelliJ plugin. It combines code gener-ation through Junie CLI with explanation generation, multiple detail levels, struc-tured and unstructured formats, explanation-to-code mappings, editor highlighting, and optional insertion of mapped comments. The methodology includes two evalua-tion directions. First, a mini-dataset based on annotated labml.ai implementations compares generated explanations with human-written explanations using an LLM-as-a-judge score. Second, a paired SWE-bench Verified Mini experiment compares normal and explanation-enriched repositories. The findings show that the plugin can produce useful explanation artifacts for realistic generation and repair workflows. In the explanation-quality evaluation, the best OpenAI configuration reached a mean score of 0.786, while the Junie-based setting reached 0.752. In SWE-bench Verified Mini, explanation-enriched repositories solved 33 out of 50 tasks, compared with 32 out of 50 in the regular condition. The results do not prove a statistically significant improvement, but they indicate that grounded explanations can support review without harming agent performance. The dissertation concludes that explanation-to-code mapping is most valuable as a verification layer for AI-generated code, while broader claims about trust and productivity require larger benchmarks and human studies.