AIvaluate: A Multi-Agent Framework for Automated Answer Script Evaluation Using Large Language Models and Semantic Vector Indexing

In conventional educational environments, it is labor-intensive, subjective, and susceptible to human error to hand-mark descriptive answers. This article introduces AIvaluate, a modular, rubric-guided multi-agent system for theory-based student answer script marking automation. The system combines Optical Character Recognition (OCR) using Azure Document Intelligence with a Cleaning Agent powered by a Large Language Model (LLM) hosted on Azure AI Foundry that cleans the text obtained without losing semantic information and ensures anonymization. Answers are parsed using RegEx to de-sentence question identifiers and mapped into corresponding correct answers and faculty-defined marking schemes in a MongoDB database. The system supports structured input prompts, where rubrics are dynamically injected into the evaluation context for consistency and flexibility across subjects. An Evaluation Agent, also driven by the LLM, matches the sanitized student answer to the correct answer and provides both a score and one-line feedback. To supplement this generative scoring, a semantic similarity-based scoring is employed via FAISS and the all-MiniLM-L6-v2 transformer model; however, this vector-based pipeline is used purely for research benchmarking and does not contribute to the final marks. Experimental outcomes indicate strong correlation between AI-computed scores and human faculty ratings, particularly in the LLM-based pipeline. This system is designed particularly for technical and theory-based academic tests, and not for literary or creative questions where language usage is the prevailing criterion. The suggested approach reduces evaluation time by approximately 70–80% compared to manual correction, without compromising fairness and transparency.