AI-Powered Tutoring System for Analyzing Novice Programmer Code

Introductory programming courses are a critical bottleneck in computer science education, often characterized by high attrition rates. A primary contributing factor is the struggle novice programmers face in diagnosing and rectifying their coding errors. They grapple not only with syntactic mistakes, which are easily caught by compilers, but more perniciously with logical and conceptual errors that prevent their programs from working as intended. Existing automated tools offer limited assistance in this regard. Compilers and linters provide cryptic error messages focused on syntax, while static analysis tools are typically designed for experienced developers and are too complex for beginners. This creates a significant gap between the student's problem and the available automated support, forcing an over-reliance on human instructors and teaching assistants for meaningful feedback. The scalability of human-provided feedback is a severe limitation. In large-enrollment courses, the instructor-to-student ratio makes it impossible to provide personalized, in-depth code reviews for every student on every assignment in a timely manner. The feedback cycle can stretch for days or even weeks, by which time the student has moved on to new topics, diminishing the pedagogical value of the feedback. This delay can lead to feelings of frustration, isolation, and inadequacy, causing students to become disengaged and ultimately drop out. Furthermore, the quality of feedback can be inconsistent, varying by the teaching assistant's experience and available time. This lack of immediate, high-quality, and consistent guidance is a major impediment to learning and a primary source of negative experiences in early CS education. This project directly addresses the challenge of providing scalable, high-quality educational support for novice programmers. The core problem is the absence of an automated system capable of understanding the *intent* behind a student's code, identifying logical flaws, and articulating constructive, pedagogical feedback. Without such a tool, educational institutions cannot effectively scale their introductory CS programs without compromising the quality of instruction and student support. Educators are left with an unmanageable workload, and students are deprived of the crucial, iterative feedback loop that is essential for mastering the complex skill of programming. The proposed solution aims to fill this gap by creating an AI tutor that acts as a patient, ever-available expert, guiding students through their errors and fostering a more effective and encouraging learning environment.

The proposed solution is a comprehensive AI-Powered Tutoring System that integrates seamlessly into the educational workflow to provide novice programmers with immediate, personalized, and actionable feedback on their code. The system's architecture is centered around a sophisticated deep learning pipeline designed to analyze student code submissions in real-time. When a student submits their code through a web-based interface, the system initiates a multi-stage analysis process. First, the code is passed to a sandboxed execution environment to capture runtime errors and output, providing a baseline of its behavior. Concurrently, the source code is parsed into an Abstract Syntax Tree (AST), which provides a structured, language-agnostic representation of the code's logic and flow. This AST is the primary input for our core machine learning model. The heart of the system is a hybrid deep learning model. We propose investigating a Graph Neural Network (GNN) that operates directly on the AST, allowing it to learn structural patterns and relationships indicative of common logical errors, such as infinite loops, incorrect conditional logic, or off-by-one errors. This GNN will be trained on a large dataset of student code submissions, each labeled with specific error types. The output of this model will be a classification of potential errors present in the code. This error classification then feeds into a second model, a fine-tuned generative language model (like a T5 or BART variant), which is conditioned on the source code and the identified error type. This second model is responsible for generating human-readable, pedagogical feedback. The feedback will not only state the error but also explain the underlying misconception and provide a hint or suggest a concept for the student to review, thereby fostering deeper learning. To manage the project's complexity and mitigate risks, development will follow an iterative approach. The initial phase will focus on building the data pipeline, including the AST parser and a system for collecting and annotating an initial dataset of Python code from university partners. The subsequent phase will concentrate on developing and training the GNN-based error classifier. Stakeholders, including students and instructors, will be involved throughout the process through user testing sessions to ensure the feedback is genuinely helpful and the user interface is intuitive. Risks such as model bias (e.g., overfitting to specific assignment types) will be addressed by ensuring the training data is diverse and representative of a wide range of problems. A feedback mechanism will allow users to rate the quality of the AI's suggestions, creating a continuous learning loop to refine the models over time. The system will be designed as a scalable web service, with a RESTful API to allow for potential integration with existing Learning Management Systems (LMS) like Canvas or Moodle, maximizing its accessibility and impact.