Fermin Romero and Marty Sullivan unveiled Cornell’s AI Platform in Las Vegas at AWS re:Invent 2025 in early December. Presented as a blueprint for universities seeking to harness artificial intelligence without sacrificing security or agility, the initiative also democratizes AI development. 

Built for moderate risk data, the platform gives end-users the tools to create their own AI agents in a secure environment, supported by a cross-unit team. For high-risk data, collaboration with Cornell’s IT Security Office makes it possible—each use case simply requires an ITSO review to ensure compliance and protection. Reference data class definitions on page 9 of University Policy 5.10 on Information Security.

Equal Access and Individual Use

"Too often, organizations designate one or two people as “the AI people” who experiment on everyone else's behalf," said Romero, a tech lead on four of the internal AI projects. "That's a bottleneck. Our gateway lets us roll out access to the entire team, whether they want to chat with a model, try the latest release from Anthropic or OpenAI, or write code against it. This isn't one person forming an opinion. It's the entire team experimenting with something new."

Romero, the Application Development Manager for Student Services IT, volunteers as a tech lead in the AI Innovation Lab. That dual role paid off when Jonathan Hart, a finance colleague who works closely with the reimbursement process, came to him with a problem. Hart described the manual grind of auditing 10,000 student group requests per semester. Romero recognized an AI use case, pitched it to the lab, and assembled a Fall 2025 team to experiment with a solution.

Building Business and Academic Solutions

"It's been a lot of long hours on top of my regular job," he said. "But after our AWS presentation, I spent hours in the hallway talking with folks from other universities and companies. A lot of them hadn't even thought about building an internal AI platform. When I explained the reimbursement automation project, that we're running real workflows in production, people's mouths hung open. One CTO told me he could only dream of that rate of adoption. When you see that reaction, you know the extra effort was worth it."

Equally impressive were the two academic use cases presented by Sullivan. As both the principal architect of CIT's AI Platform and a PhD student in the Graduate School, Sullivan sits squarely in the teaching, learning, and research environment. Like Romero, he dove into the opportunity to utilize AI when his immediate customers sought solutions to improve their teaching, advising, and coaching activities.

Two of the use cases Sullivan presented at the conference were built to address different academic challenges: providing effective coaching for first-year engineering students and creating a Socratic learning environment to ensure students in large classes got individual guidance and challenges to assess their understanding in a non-threatening manner.

In both cases, Sullivan leaned on graduate students to help develop and test the solutions.

The Bigger Picture: Creating Builders

"Cornell isn't just solving problems. It's creating builders and entrepreneurs," said Sullivan. "That's who we are at Cornell."

The 70 graduate students earning independent study credit aren't working in isolation. Each team is paired with a tech lead from Cornell's IT organization, professionals who build and deploy enterprise grade solutions every day. That mentorship bridges the gap between academic projects and production ready systems.

For students, the payoff is tangible: a working demo of software currently in use, not just a line on a CV. The platform supports different users in different ways. Developers can write code directly with code assistants like Claude Code safely by utilizing the models through the AI Gateway or connect it to agentic tools. Graduate students build automations using N8N, a visual workflow platform that will power the upcoming AI Agent Studio. Faculty and staff can experiment with models through a chat interface, no coding required. Together, these layers foster a culture of innovation and collaboration across the university.

Campus staff and faculty interested in using the AI Platform Service components, like the AI Gateway or the future AI Agent Studio, can reach out to the AI team through the IT Service Desk to set up their team.

Three Use Cases Illustrate the Platform’s Impact

The team selects projects based on four criteria: impact, feasibility, alignment with institutional priorities, and ethical responsibility. An engaged project sponsor is essential. The following three use cases represent recent successes.

Values Exploration and Reflection Assistant (VERA)

The Selander Center for Engineering Leadership needed a scalable way to help 900 first-year engineering students reflect on their individual values and goals. Assistant Dean Erica Dawson, the center’s director, built her system prompt using Claude for Desktop. She organized the grounding research data and Sullivan plugged the data and prompt into an agent with guardrails. During its July/August pre-semester launch, VERA logged over 40,000 student interactions with zero support requests after initial configuration.

“VERA resulted in accelerated values discovery and much richer live coaching sessions because both the student and the coach could enter the conversation from a place of deeper understanding,” said Dawson.

Socratic Chatbot for Classroom Learning

In large science courses, instructors wanted students to think critically, not just answer questions. Sullivan worked with his academic advisor Toby Ault, Associate Professor & Director of Graduate Studies in the Earth & Atmospheric Sciences department. Ault teaches Climate & Energy, a course with around 300 students. Sullivan worked closely with Amazon Web Services and Ault to integrate a custom Socratic Chat application into Canvas LMS, leveraging Bloom's Taxonomy and adding progress bars for transparency and to gamify the classroom activity. Students now engage with topics as they would in office hours, while instructors gain analytics on comprehension. 

Read their full story about transforming classroom conversations with AI-powered Socratic Chat on the AWS Public Sector blog.

Campus Groups Audit Automation

Processing 10,000 reimbursement requests per semester for 1,600 student groups had grown into a months-long manual chore. Working closely with the project sponsor, the tech leads and graduate student developers built an AI workflow to validate receipts and compliance rules, saving 30+ minutes per submission. Every automated transaction is still human-reviewed, but confidence is growing as accuracy scales.

Lessons Learned

"One question that came up constantly in my hallway conversations at the conference: where do we even start? Based on our completed projects, we distilled a few principles that helped us move quickly and can help other teams do the same," said Romero.

• Define the problem before building. Be sure to start with a clear use case that focuses on a specific problem, not a platform.
• Work with the people who personally experience the problem. Managers may ask for a solution for their team, but identifying the pain points from the ground up makes it easier to successfully create a lasting solution.
• Iteration is key. A proof of concept is quick to develop, so ship something, learn, and improve.
• Human in the loop vs. on the loop. Choose the level of human review based on risk. Some solutions need every output checked; others can rely on spot checks. This can shift over time as confidence grows.

Watch the full presentation, recorded at AWS re:Invent 2025 and published on YouTube.