Implementation of the AISE Project AI methodology
Four state-of-the-art Large Language Models (LLMs), including the GPT-4o-mini, GPT-4.1-mini, DeepSeek-Chat, and DeepSeek-Coder, were systematically benchmarked within a robust Retrieval-Augmented Generation (RAG) architecture for automated essay grading across the Content, Organization, and Language dimensions. Each LLM was evaluated using the accuracy, macro-precision, macro-recall, macro-F1, balanced accuracy, and confusion matrix analysis on the held-out test set of 261 essays. The RAG-based GPT-4o-mini emerged as the clearly superior model across all three dimensions, achieving accuracy of 0.820 (Content), 0.790 (Organization), and 0.740 (Language), with macro-F1 scores of 0.813, 0.774, and 0.724 respectively, substantially outperforming the other candidate models. It also demonstrated the lowest and most stable inference latency (1.7–2.1 seconds per essay), confirming its suitability for real-time deployment. Based on these results, GPT-4o-mini was selected as the production model for the AISE grading engine. The final system was implemented as a RAG-enabled FastAPI service, exposing a /grade endpoint that accepts essay text and authentication credentials and returns dimension-specific grades in structured JSON format. The service is deployed on a dedicated university server and fully integrated with the AISE educational platform to enable real-time, multilingual, and pedagogically grounded automated essay assessment for participating schools.