Large Language Models (LLMs) are increasingly integrated into computing tasks, from chatbots to code generation, yet we often treat them as opaque tools rather than programmable systems. We present a lightweight, reproducible platform that enables students to experiment with LLMs in a private, web-based environment. By combining GitHub Codespaces with the open-source inference en- gine llama.cpp and small quantized models, the platform removes complex installation barriers and reliance on commercial services, enabling scalable, cost-effective experimentation. Students can ex- plore prompt engineering, memory, and stochasticity while directly inspecting model behavior, treating LLMs as first-class program- ming primitives rather than black-box tools. The workflow inte- grates seamlessly with GitHub Classroom, reducing setup time and administrative overhead while providing access to higher-powered resources. In a pilot study, students engaged deeply with LLM be- havior, uncovering limitations in smaller models and developing critical perspectives on AI as context-dependent computation rather than sentient intelligence. This approach offers a unified frame- work for diverse courses, from CS1 to computing literacy courses, fostering both technical skills and critical AI literacy.

Jason Madar is a computing educator and practitioner with over 25 years of experience bridging industry, research, and classroom practice. His current work focuses on hands-on AI literacy, integrating large language models into introductory computing courses, and developing accessible, reproducible platforms for exploring AI systems. Drawing on a background in data mining and decades of industry experience, Jason designs practical workflows and innovative teaching strategies that make AI tangible for students, while emphasizing critical thinking, privacy, and computational understanding.