Software engineering programs have long served to prepare students to apply rigorous software engineering practices in the industry, but the rapid adoption of LLMs in development workflows has outpaced curricula. The problem is not that students are using LLMs; they already are. The problem is that they are using them without a disciplined engineering model for doing so, exposing themselves and their projects to well-documented risks: incorrect outputs, security vulnerabilities, accumulating technical debt, and the erosion of the very supervision skills that make human oversight of AI-generated artifacts possible.
This talk presents a metamodel framework for LLM-assisted software engineering, developed for a graduate MSE capstone program as a practical instrument for enabling students to treat their AI-assisted software engineering activities as a designable sociotechnical system rather than a collection of ad-hoc tools. Drawing on the factory analogy where workstations, quality control gates, production metrics, and material handling must be deliberately configured, the framework gives students a vocabulary and a design process for making explicit decisions about what LLMs produce, how outputs are validated, who holds approval authority, and how quality is measured. I will discuss the ideas behind the four-layer metamodel (Artifact, Process, Resource, Measurement), a five-step system design process that faculty can assign as a capstone deliverable, and a risk taxonomy covering both output quality and organizational failure modes.
School of Computer Science
Carnegie Mellon University
Scott Pavetti is an Assistant Teaching Professor in Carnegie Mellon University's Master of Software Engineering program, where he has taught core graduate courses including Software Architecture, Agile Methods, Engineering Embedded Systems, Requirements for Embedded Systems, and Statistics for Decision Making, among others, since 2020. He serves as a Studio and Practicum Project Mentor for industry capstone projects. His research contributions include a peer-reviewed publication on DevSecOps pipeline security (Woody et al., 2020, Journal of Systemics Cybernetics and Informatics) and a co-authored SEI technical report applying Model-Based Systems Engineering to DevSecOps pipeline assurance (Chick, Pavetti, and Shevchenko, 2023). Most recently, he has been developing a practitioner framework for structuring LLM integration into the software development lifecycle.
Prior to joining the faculty, Scott wore a number of job titles, such as research programmer, senior software engineer, tech lead, and software quality engineer, accumulating a broad set of product development experiences ranging from mobile and desktop development, small IT deployments, embedded systems, and IoT systems. He spent two years as a Community of Practice Leader in Software, where he worked with 50 engineers in the US and Germany to develop better practices and share knowledge, and has trained engineers in software architecture, design, requirements, agile, and testing practices. He understands that being a software engineer takes more than technical ability.
He got his undergraduate degree in Computer Engineering from the University of Pittsburgh and a master's degree in Software Engineering from Carnegie Mellon University.
