Antonio Jorge Forte, M.D., Ph.D.
SafeGuardED: A RAG-Based AI Assistant for Trusted Resident Training
Building Safer, Specialty-Grounded AI for Surgical Education
Award year: 2025 Richard M. Schulze Scholar in Artificial Intelligence
My hope is that trainees will have access to a tool that supports learning in a safe and structured way. Surgical residents are constantly preparing for cases and reviewing complex topics. If they can ask questions and receive answers grounded in trusted specialty sources, that can reinforce high-quality standards.
You’ve spent years working at the intersection of surgery, research and education. What sparked your interest in developing this AI tool for trainees?
For the past six or seven years, our Mayo Clinic Surgery AI Lab has been working with natural language processing — well before large language models became widely known. We’ve always been interested in how technology can support both clinical practice and education.
When large language models emerged, we immediately saw their potential. Trainees are constantly asking questions as they learn, and many are already using public AI tools. The challenge is that those systems generate answers based on internet data. Some of that information is excellent. Much of it is not. In medicine, that variability is not acceptable.
Rather than ignore that reality, we asked: How can we build a version that is safe, grounded in high-quality specialty content, and truly designed for medical education? That question is what sparked this project.
In simple terms, what are you building?
We are building a specialty-specific AI agent for surgical trainees, beginning with plastic surgery.
The foundation is a large language model, but the key difference is how it generates answers. Instead of relying on general internet data, we ground the model in vetted, specialty-specific textbooks and curated educational materials. This approach is called retrieval-augmented generation.
In practical terms, when a resident asks a question, the system retrieves relevant information from trusted plastic surgery sources and generates an answer based only on that content. It’s almost like the model opens the specialty textbook first and then responds rather than answering from broad, unfiltered data.
The tool will be accessible through a secure Mayo Clinic web address, allowing residents to use it from a computer or mobile device in real time.
Large language models can be unpredictable. How are you ensuring safety and rigor?
Safety and benchmarking are central to our work.
Not all retrieval-augmented generation systems are built the same way. There are different strategies for retrieving and ranking information before it reaches the model. Part of this project is evaluating those strategies to determine which approach produces the most accurate and reliable responses.
We are currently curating the specialty content that will ground the model. Once integrated, we will pilot the system with residents under appropriate research oversight.
We are not simply building a chatbot. We are rigorously evaluating whether it improves learning and whether its outputs meet specialty standards. In medical education, innovation must always be paired with validation.
What impact do you hope this work will have?
My hope is that trainees will have access to a tool that supports learning in a safe and structured way. Surgical residents are constantly preparing for cases and reviewing complex topics. If they can ask questions and receive answers grounded in trusted specialty sources, that can reinforce high-quality standards.
Long term, this framework could extend beyond plastic surgery into other specialties. The broader goal is to elevate how AI is used in medical education — moving from general-purpose tools to carefully designed, specialty-grounded agentic systems.
How has philanthropic support shaped this work?
The support as a Richard M. Schulze Scholar in Artificial Intelligence has provided critical momentum. We had conceptualized the model, but we lacked the resources to fully curate the content, build the grounded system and begin implementation.
Philanthropic support allowed us to move from concept to execution — with an emphasis on safety, quality and responsible use. It has enabled us to approach AI thoughtfully rather than reactively, aligned with Mayo Clinic’s educational standards.