PockeTA: Virtual Teaching Assistant Platform with Course Context

Teaching assistants (TAs) are essential for scalable, high‑quality university education, yet most institutions face persistent constraints in TA-to-student ratios, limiting timely feedback and personalized support. At the same time, large language models (LLMs) are rapidly being adopted in education, but typical deployments treat them as generic chatbots, lacking course awareness, pedagogical grounding, and meaningful integration with instructors’ workflows. In this work, we present PockeTA, an ubiquitously accessible virtual teaching assistant platform that tightly couples LLM-based multi-agent intelligence with a teacher-controlled, graph-based knowledge base (K-Cube). PockeTA provides a personified avatar interface on both desktop and mobile, supports multimodal interaction (text, document sharing, interactive drawing), and is deeply aware of course structure and materials through K-Cube integration. Student interactions are continuously collected, organized as learning analytics, and visualized to help instructors monitor learning progress, identify misconceptions, and refine teaching. We describe the overall system design, including its multi-agent architecture and knowledge-graph integration, and report initial findings from a beta deployment with university students. Early user feedback indicates that PockeTA complements human TAs by providing always-available, context-aware support, while giving teachers unprecedented visibility and control over the virtual assistant’s behavior and course context. Our results suggest that PockeTA offers a promising path toward scalable, course-grounded virtual TAs that go beyond generic LLM interfaces to support both learners and educators.