Code Generation and Debugging Assistant using Deep Learning and Large Language Models

Abstract: Novice programmers frequently struggle to translate cryptic compiler diagnostics into actionable understanding, while contemporary generative Artificial Intelligence (AI) assistants optimized for professional productivity tend to resolve reported errors by supplying complete corrected code immediately. This mode of instant resolution, though efficient for experienced developers, is pedagogically counterproductive for learners, as it discourages active engagement with the debugging process. This paper presents CodeMentor AI, a full-stack hybrid web application combining a locally hosted statistical Machine Learning (ML) pipeline with a cloud based Large Language Model (LLM) to deliver both rapid error triage and structured, Socratic-style tutoring. The FastAPI backend integrates a TF-IDF vectorizer with a Logistic Regression classifier for baseline error category prediction, and a cosine-similarity Nearest Neighbors model for code retrieval, with the Groq-hosted openai/gpt oss 120b LLM used for explanation, progressive hinting, and solution verification.The React frontend provides an integrated Monaco-based workspace, a three-tier Learning Mode hint system, and an administrative analytics dashboard with Firebase Authentication and a Supabase (PostgreSQL) backend for persistence. This paper documents the system's architecture, module design, and AI pipeline as implemented, and situates it within recent literature on AI-assisted programming education, retrieval-augmented debugging, and comparative studies of commercial AI coding tools. Formal empirical evaluation of learning outcomes and system accuracy has not yet been conducted and is identified as future work.

Keywords: Artificial Intelligence in Education, Large Language Models, Retrieval Augmented Generation, Programming Tutoring Systems, Automated Debugging, Hybrid Machine Learning, FastAPI, React, Supabase, Socratic Scaffolding