WIP: Tutoring System Using a Large Language Model

This innovative practice Work-in-Progress (WIP) paper describes early design, implementation, and pilot evaluation of SSE-EduBot, a course-specific tutoring assistant that integrates semantic search with a LoRA-fine-tuned Llama-213B model. AI in education is a promising domain for the utilization of LLMs to make personalized learning materials using elements of generative AI. This can help educators 'riff' on traditional teaching by leveraging Gen AI as tutoring systems. We design a system targeted for a 100 + student graduate data science course. The three-stage pipeline (1) retrieves answers from prior instructor-student emails, (2) invokes the fine-tuned model when no archival match is found, and (3) reformats the output into conversational prose for a web interface. A curated corpus of lecture transcripts, slides, and textbook excerpts, augmented with ≈ 4,300 automatically generated question-answer pairs, serves as the domain knowledge base. Preliminary comparisons against ChatGPT-3.5 on representative queries show that SSE-EduBot delivers more context-aligned explanations and reduced domainspecific hallucinations/verbose explanations. Deployment on the university server for the upcoming semester will help collect real-world interaction data to refine retrieval precision, response latency, and ethical guardrails. We share the reproducible workflow and outline next steps toward scaling the approach across additional courses and disciplines.