Parrot: An Agentic Classroom AI

Kalena Dai, Arya Sarukkai

Keywords: Agentic AI, Classroom AI, Multimodal sentiment analysis, Retrieval-Augmented Generation (RAG), Adaptive learning, Federated learning, Privacy-preserving AI, Educational technology, Large Language Models (LLMs), Explainable AI

TL;DR: Parrot is a privacy-preserving, multimodal AI agent that enhances classroom learning by summarizing, questioning, and adapting through federated collaboration and interpretable feedback.

Abstract: We introduce Parrot, an interpretable, multimodal AI agent designed to enhance real-time teaching and learning in classrooms. Parrot operates autonomously as both a curious student and an assistant lecturer, performing actions such as summarizing lecture content, detecting engagement via multimodal sentiment analysis, and generating context-aware questions. The system integrates Retrieval-Augmented Generation (RAG) grounded in curriculum materials, DeepPrivacy2 for real-time face anonymization, and adaptive learning capabilities. Each classroom instance locally adapts its strategies while contributing anonymized metadata to improve shared retrieval and prompt policies via federated collaboration. A dedicated Learner module continuously refines Parrot’s retrieval logic and prompting behaviors, enabling long-term improvement without compromising privacy. We present results from simulated deployments and discuss how Parrot exemplifies agentic intelligence in education through adaptability, transparency, and trustworthy autonomy.

Submission Number: 12