Middleware-enabled LLM feedback for automated grading platforms

Details

Supervisors

Mens, Kim ; Nijssen, Siegfried

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

enThe rapid diffusion of Large Language Model (LLM) technology is reshaping educational software, yet most Automated Grading Platforms (AGPs) still restrict feedback to binary verdicts or static hints. This lack of personalised, context-aware guidance curtails their pedagogical impact, especially in introductory programming courses where formative support is critical. Many students resort to uncontrolled use of external AI assistants, which undermines the feedback loop and threatens academic integrity. This thesis investigates how AI-generated feedback can be integrated into existing AGPs without entangling platform code with fast-evolving AI services. We propose a provider-agnostic middleware that (i) decouples platform-specific details through modular adapters, (ii) enables dynamic model selection, and (iii) supplies a prompt-template system that could embed unit-test outcomes, code context, and instructor goals into reproducible LLM prompts. A reference implementation (built with Flask, MongoDB, and Docker) exposes a REST API and a lightweight Python client, enabling educators to adopt the solution with minimal effort and to experiment with alternative models or prompt strategies. Effectiveness was evaluated in two validation phases within a first-year computer-science course on INGInious. Across the two phases, AI feedback increased the immediate success rate on the next attempt by +5.02 and +15.10 percentage points, almost all students (94.4% in Survey 1 and 100% in Survey 2) agreed that the tool can be an asset for improving programming learning, and over 75% indicated that such a tool could reduce their use of external AI assistance. These findings show that the proposed middleware transforms static AGP verdicts into targeted formative feedback that improves learning outcomes, heightens student satisfaction, and curbs unregulated AI usage, without sacrificing platform maintainability. The main contributions are: (1) a reusable, provider-agnostic middleware architecture; (2) an open-source implementation; and (3) a mixedmethod validation demonstrating statistically significant quantitative gains and rich qualitative insights.