In today’s educational landscape, generative AI tools have become both a blessing and a challenge. While these tools offer unprecedented access to information, they’ve also created new concerns about academic integrity. Increasingly, students rely on AI to generate direct answers to homework questions, often at the expense of developing critical thinking skills and mastering core concepts.
Universities and educators face the growing dilemma of how to harness the power of AI as a teaching tool while ensuring it supplements rather than short-circuits the learning process. The key lies in developing systems that guide students toward understanding rather than simply handing over answers.
Alongside concerns about academic integrity, many universities are also grappling with a shortage of teaching assistants (TAs). At Clemson University and institutions across the country, rising enrollment in computer science and other STEM fields is outpacing the availability of TA support. In some courses, the student-to-TA ratio is 20 to 1. During busy periods such as midterms and project deadlines, this imbalance often leaves students without the timely help they need.
To address this challenge, our Clemson capstone team set out to build a virtual teaching assistant (VTA) using the NVIDIA AI Blueprint for retrieval-augmented generation (RAG), supported by NVIDIA NeMo Retriever and NVIDIA NIM microservices. Our goal was to develop a scalable, always-available learning assistant that encourages student learning by guiding them through concepts and problem-solving, rather than simply providing answers.
Our solution: a concept-focused virtual teaching assistant
Our team, in collaboration with NVIDIA, developed an AI-powered teaching assistant built on the NVIDIA RAG Blueprint designed to explain underlying concepts, guide students through problem-solving processes, and ultimately foster deeper understanding.
What makes our approach unique is the integration of three key components:
- NVIDIA RAG Blueprint: Supplies fast, retrieval-augmented generation that pulls only the most relevant course fragments, anchors every explanation in verifiable sources, and formats citations automatically. This backbone lets the assistant stay fact-based while still sounding conversational.
- NVIDIA NeMo Guardrails: Adds ethical guardrails that intercept “just give me the solution” queries and convert them into scaffolded guidance. Custom actions classify intent, reformulate prompts, and enforce citation checks, ensuring the AI encourages learning and complies with academic integrity policies.
- Course data manager API: Bridges Canvas with the RAG knowledge base, syncing modules, deadlines, quizzes, and uploaded files. With this live context, the assistant tailors examples, reminders, and hints to the exact material each student is studying, making responses feel course-specific rather than generic.
Technical implementation
Our AI-powered TA system leverages multiple docker containers from NVIDIA and the Canvas platform UI.
Building on the NVIDIA AI Blueprint for RAG
Our implementation uses NVIDIA AI Blueprint for RAG as its foundation. This architecture enables our teaching assistants to:
- Process and understand student queries contextually.
- Retrieve relevant course materials and resources.
- Generate responses that focus on conceptual understanding rather than direct answers.
- Maintain conversational context across interactions.
NeMo Guardrails: ensuring ethical AI use in education
The need for citation verification has become more pressing with the increasing adoption of large language models (LLMs). Recent advances in RAG methods help reduce hallucinations in generated content. However, significant challenges remain in establishing trustworthiness without additional verification methods.
A critical improvement in our system is the integration of NeMo Guardrails, which:
- Analyze student questions to determine if they’re seeking conceptual understanding or direct answers.
- Redirect direct homework question queries toward conceptual explanations.
- Provide scaffolded guidance that leads students to discover answers themselves.
Custom actions in NeMo Guardrails
Actions are small Python functions that NeMo Guardrails can invoke at specific points in the conversation to inspect, modify, or replace the model’s reply. Each action sees the full chat context and can call APIs, filter text, or generate a new message before handing control back to the workflow. This lets our VTA layer learning-centered prompts and hard citation checks on top of a RAG pipeline, meeting the bar for verified, trustworthy LLM answers without retraining the base model.
We added three such actions:
Quiz_response: Detects quiz-style questions and rewrites the prompt so the assistant explains the core concept, offers a brief example, and ends with a guiding question.Homework_brainstorm: Turns homework queries into a structured series of hints that break the task into steps while withholding the final answer.Code_debug: Walks students through debugging strategies, pointing out likely error spots and suggesting tests, without fixing the code outright. Together, these actions prevent direct-answer leakage, ensure all responses remain grounded in top-k-retrieved course materials with inline citations, and promote active learning and comprehension.
These are just a few examples of what’s possible, as the system is highly configurable and can be extended by end users to support any instructional or organizational needs.
Course data manager API
We developed a custom course data manager API that integrates with Canvas through secure access tokens. This integration enables:
- Awareness of course content, assignment deadlines, and context. By automatically syncing the latest modules, files, and due dates, the assistant can reference current topics and remind students of upcoming milestones from Canvas.
- Personalized responses based on specific course materials. Because every retrieved item is indexed alongside its course metadata, the assistant tailors explanations, hints, and citations to the exact readings, lectures, or rubrics each student is working with, creating guidance that feels course-specific rather than generic.
User interface and experience
Our virtual teaching assistant presents an intuitive interface that encourages learning-focused interactions:
- A conversational interface that supports follow-up questions and preserves chat history across turns so students can refine or dig deeper without losing context.
- Clear indication of when conceptual guidance is being provided. This is highlighted by subtle labels and color cues that signal the assistant is coaching rather than giving solutions.
- An interface for getting course information. A dedicated panel connects the virtual teaching assistant to class materials and files, so it can pull in the right context on demand.
Impact on education
The implementation of our AI teaching assistant offers several key benefits to educational institutions:
- Enhanced learning: By guiding students through underlying principles instead of supplying solutions, the assistant strengthens conceptual mastery, encourages metacognition, and prompts learners to articulate their own reasoning. This active process mirrors a one-on-one tutoring session, reinforcing long-term retention and problem-solving confidence.
- Academic integrity: Guardrails that block direct-answer extraction, require source-backed explanations, and log questionable requests help discourage shortcut-seeking and plagiarism. Instructors can trust that AI support aligns with course objectives and assessment policies rather than undermining them.
- Scalable support: The assistant delivers individualized, context-aware help around the clock, something impractical for human TAs at large enrollment levels. As usage scales, each student still receives timely feedback tailored to their course materials, freeing faculty to focus on higher-value teaching activities.
Future developments
While our current implementation provides robust support for concept-focused learning, we see some avenues for future enhancement. A campus-wide pilot will be implemented at Clemson in partnership with the university’s central IT group to provide general student access. On the technical side, guardrails can be strengthened to prevent actions that make increasingly subtle attempts to elicit full solutions, ensuring the tutor remains learning-centered. An accompanying analytics dashboard will be added to report system failures, including common misconceptions, unanswered questions, and citation gaps, so instructors can refine course materials. Non-code extension hooks will be developed to enable other institutions to easily customize and integrate the Blueprint with their Canvas account or other learning management platforms.
Conclusion
The VTA we’ve developed demonstrates how AI can enhance rather than undermine educational objectives when properly implemented. By focusing on concept explanation rather than answer provision, our system helps preserve the integrity of the learning process while still offering valuable support to students.
As educational institutions continue to navigate the complexities of AI in academic settings, solutions like ours provide a blueprint for responsible implementation that upholds educational values while embracing technological advancement.
Looking forward, other schools can extend this blueprint by swapping in their course content, fine-tuning guardrails for local policies, and packaging discipline-specific knowledge packs. Because the codebase is modular and already integrated with Canvas, institutions can quickly pilot the assistant and, through the blueprint’s release, adapt it to additional LMS platforms, add custom analytics, and share new features with the community, fostering a collaborative ecosystem of AI-powered learning tools.
To access our codebase and run our application, visit GitHub.
Acknowledgments
This project was part of the Spring 2025 Clemson University Capstone Program, supported by the Clemson School of Computing and Department of Computer Science. The name and idea, Dori, were inspired by the Fall 2024 Capstone team: Chloe Crozier, Ethan Anderson, Justin Silva, and Teja Guruvelli. Special thanks to Professor Carrie Russell and our faculty reviewers and peers for their valuable guidance and reviews.
