Introduction
Senior capstone projects serve as the critical bridge between undergraduate academic theory and professional engineering practice. However, integrating Artificial Intelligence (AI) into these projects presents unique pedagogical challenges. AI is often perceived by students as a "black-box" domain, characterized by "fuzzy" requirements, complex hardware integration, and the difficulty of establishing ground truth for model training.
This paper presents a structured framework for managing complex AI projects within a two-semester undergraduate computer science curriculum. By blending disciplined planning with iterative development, the authors demonstrate how students can successfully build high-impact systems, ranging from autonomous assistive technologies to machine-learning-driven security tools.
The Instructional Framework
The authors propose a 25-week instructional approach designed to mitigate the risks of long-term AI development. This timeline is divided into four distinct phases:
- Initiation (2 weeks): Students are matched to projects based on their technical preferences and prior experience in AI or related fields.
- Planning (3 weeks): A "mini-Waterfall" phase where teams define the project scope, requirements, and architectural design before writing code. This upfront planning is essential for AI projects to prevent teams from becoming overwhelmed by the vast possibilities of implementation.
- Development (20 weeks): An Agile/Scrum approach consisting of 2–3 week sprints. Each sprint includes planning, implementation, and a demonstration of progress to mentors.
- Termination (3 weeks): The final phase focuses on comprehensive reporting and a public poster or demonstration session.
Technical Implementation and Case Studies
The framework was validated through several diverse projects that utilized a variety of AI methodologies and data sources:
Autonomous Navigation for ALS Patients
Students developed a semi-autonomous wheelchair designed for users with Amyotrophic Lateral Sclerosis (ALS). This project utilized Microsoft Kinect sensors to process real-time RGB, depth, and infrared (IR) data. The team compared four doorway detection methods, ultimately finding that an IR-based detector using reflective tape was significantly more robust than Haar cascade classifiers, which suffered from high false-positive rates in real-world environments.
Network Security and Anomaly Detection
In the domain of cybersecurity, a team utilized the ExtraHop platform to acquire raw packet data. This data was stored in NoSQL databases and used to train Support Vector Machines (SVMs) for network anomaly detection. A primary challenge for this team was the "ground truth hurdle"—the difficulty of accurately labeling "suspicious" versus "normal" traffic without a pre-defined dataset.
Educational Data Mining
Another project focused on analyzing player behavior within a gamified learning platform. By collecting statistics on homework submissions, grades, and attendance from a large introductory CS class, students attempted to predict engagement. While MBTI personality types showed little correlation with usage, the team found that specific gaming typologies (such as BrainHex) provided a promising avenue for predicting student motivation.
Key Insights and Pedagogical Lessons
The Agile-Waterfall Hybrid
While Agile development is the industry standard, the authors found that a purely iterative approach can lead to "scope creep" in AI projects. The inclusion of a short, upfront Waterfall-style planning phase proved crucial. It forced students to establish a defined goal and baseline before diving into the complexities of model training and hardware integration.
Motivation through Social Utility
Projects with a clear "human" benefit—specifically the assistive wheelchair—demonstrated higher levels of student persistence. The social utility of the project served as a powerful motivator, helping students push through the technical frustrations inherent in robotics and computer vision.
Iterative Complexity
The authors emphasize the importance of starting with simple, "dumb" algorithms to establish a baseline. By building a working system early, teams could then iterate toward more complex AI models with a clear understanding of the performance improvements required.
Limitations and Future Directions
Despite the success of the framework, the authors noted significant instructor overhead. AI projects require more technical guidance and active mentorship than standard software engineering projects. Additionally, hardware-centric projects face a "hardware bottleneck," where software debugging is impossible without a stable physical platform. Future work will focus on refining data mining techniques to better predict student engagement across different demographics, potentially contributing to higher-level research in educational technology.
Summary updated