Content area
This thesis explores the development and empirical evaluation of a Large Language Model (LLM)-based multi-agent AI Tutor designed to enhance student learning in the context of elevator pitch creation. The AI Tutor system was implemented using LangChain and Chainlit, integrating OpenAI's GPT-4o model to simulate four educational agent roles: Mentor, Peer, Evaluator, and Progress Tracker. Each agent provided structured, adaptive support aligned with constructivist learning principles. The system was deployed in a real-world classroom experiment at NOVA IMS with higher education students, comparing the effectiveness of the AI Tutor against traditional instruction. The study employed a quasi-experimental design, with participants divided into two groups: one using the AI Tutor application and the other receiving a conventional mini-lecture. Data collection included pre- and post-session surveys capturing perceived learning gains, engagement, and satisfaction, along with elevator pitch submissions evaluated by a jury using a standardized rubric. Results indicate that the AI Tutor group demonstrated higher levels of perceived engagement and self-reported improvement in pitch development skills, and their final submissions showed greater clarity, structure, and creativity. The findings suggest that LLM-based AI agents, when structured as collaborative tutors, can meaningfully support short-format learning in higher education. This work contributes to ongoing discussions on AI in education by providing practical insights into system design, implementation, and real-world classroom integration. Limitations and opportunities for future research are also discussed, including enhancements to long-term memory, adaptive analytics, and multi-modal capabilities.