Abstract:

Artificial Intelligence in education is not a novel innovation. In the mid-20th century, Intelligent Tutoring Systems (ITS) such as SCHOLAR and PLATO were developed to emulate educator feedback. This spawned the usage of automatic grading tools and writing feedback tools to offload the burden of low-stakes work for educators. The use of Large Language Models (LLM) in educational settings exploded in 2022, when OpenAI publicly released ChatGPT as an LLM innovation that transformed both K-12 and higher education. Currently, 89% of high school and college students avail of its services, and more than 50% of educators also reap the benefits of ChatGPT supporting Generative Artificial Intelligence (GAI) to generate course materials.

Norman, with a nod to Donald Norman, a founder of user-centered design and usability engineering, is a multi-tenant AI teaching assistant that addresses weaknesses of existing models such as hallucinated citations, absence of source transparency, and decontextualized responses. Through Retrieval-Augmented Generation (RAG), role-based access control, and traceable source attribution, Norman utilizes faculty-posted course materials in a model that supports student interactions with validated educational content, eliminating hallucinations and the opacity characteristic of general-purpose LLMs. Built in Python using a Flask application server, Norman executes inferences offline using the open-weights Qwen3.5 foundation model. It converts content into dense vector embeddings using ChromaDB as the vector store, ensuring all data processing and inference remain strictly within the university's on-premise environment.

Usability Evaluations will be conducted to provide structured findings of usage metrics, qualitative student and faculty feedback, as well as analysis of retrieval accuracy specific to course content. Norman will initially be deployed in a higher education undergraduate class to validate the technical architecture under authentic class conditions. Norman’s focus is usability of an AI-teaching assistant, generating a system image that reflects designer intent with user mental models, through a secure fluid interface that supports the needs of both higher education faculty and the student population.

Keywords:

Artificial Intelligence in education, large language models, Human-centered design.