The Boundary-Spanning Assistant: Understanding the Role and Usage patterns of LLMS in Project-Based Software Engineering

Abstract

Large Language Models (LLMs) are rapidly entering educational settings, yet their role in project-based software engineering (SE) courses remains insufficiently understood. Most discourse centers on code correctness or academic integrity, with limited attention to how LLMS support learning across project phases. This paper investigates how an LLM-based course assistant, InnSpill.ai, shaped students' learning experiences. We report on an exploratory study involving 76 first-year students organized into 15 teams, drawing on team reports, weekly usage polls, and supervisor observations. Using thematic analysis, we examine interaction patterns, perceived value, and limitations of the tool. We found that students engaged most intensively with InnSpill during early and mid-project phases, primarily for conceptual clarification, documentation structuring, and lightweight code generation. The analysis identifies six boundary-spanning roles enacted by the assistant, ranging from technical support to course interpretation. While students valued the assistant's accessibility and formative guidance, challenges included uneven adoption at the team level, tendencies toward overreliance, and the absence of conversational memory. The findings suggest that when deliberately embedded into reflective and iterative project workflows, LLM tools can enhance formative learning by providing just-in-time support and reducing cognitive barriers. Future work should examine collaborative prompting practices and multimodal extensions to better align LLM affordances with team-based SE learning.