As AI chatbots become more popular, people have increased the amount of confidential information shared with these tools — often without knowledge of the consequences. TrustLayer AI, by Summer 2026 Master of Information and Cybersecurity alums Brandon Tarca and Sash Upadhyaya, aims to read through chatbot prompts before submission, scan for sensitive information, and make decisions around content redaction in order to address this issue.
The team was awarded the Summer 2026 Lily L. Chang Capstone Award, which recognizes the semester’s top project.
To learn more, we interviewed the team —
What inspired your project? How did you decide on the concept?
Sash: During my machine learning class, I experimented with gateways and how models could be used for classification, specifically for SQL injection. I continued experimenting with different inline capabilities, and in my web application security class, while analyzing an AI portal that UC Berkeley was rolling out, I realized the same concept could be applied to AI prompts to protect sensitive information sent to an external model. I continued the work and finalized an integrated solution in March.
During my privacy engineering class, I focused on a project measuring how utility in large language model (LLM) output can be preserved while using different methods of redacting personally identifiable information (PII) in prompts. The next semester, I took Natural Language Processing (NLP) with Deep Learning and built a strong foundation in semantic analysis. This project became a natural extension of both classes, and I had a lot of interest in this topic, as PII being exposed to external LLMs has been a widespread issue in 2026.
What was the timeline or process like from concept to final project?
Sash: Since I had the gateway finished before the capstone class began, it allowed our team to experiment with different classification methods. To improve the demo, Brandon suggested that we integrate a messaging user interface (UI) that provided a familiar interface most users understand when interacting with AI. We leveraged GitHub to collaborate on the code and run local instances on our own devices. As the capstone presentation approached, we stopped adding new features, finalized the code, and deployed the solution to Amazon Web Services (AWS). Brandon’s messaging UI idea became the centralized platform for the demo, helping bring everything together into one cohesive solution.
Brandon: The process of building the product involved multiple phases with a lot of iterative testing and trimming of scope until we landed on something that worked well. Our idea at its core was ambitious — our platform needed to be accurate enough to be trusted by enterprises in sensitive industries, efficient enough to run in seconds, and permissive enough so employees do not experience unnecessary blocking. Building out our engine took a while, and we still have a lot of room for improvement, but our core architecture is strong and rooted in the research we’ve studied throughout our graduate classes.
“One of the things Brandon and I felt strongly about was tying together what we learned across the MICS program into one cohesive solution. The final project was not based on a single class or technology, but on applying concepts from across the curriculum to build a working security platform.”
How did you work as a team? How did you work together as members of an online degree program?
Sash: We worked great as a team and collaborated using Slack, with a Friday team sync throughout the capstone project. In addition, we both pursued the Graduate Data Science Certificate, which was helpful in understanding and configuring the models used within TrustLayer AI.
Brandon: Sash and I have worked together in previous classes and are both very interested in detection engineering, which I believe led to our strong team synergy.
How did your I School curriculum help prepare you for this project? Which course in the program was uniquely helpful in preparing you for your capstone project and why?
Sash: Mathematics and Programming for Cybersecurity provided the foundation for integrating the gateway with our classification models using Python. Network Security gave us the knowledge to handle DNS integration, gateway provisioning, and low-latency communication with external services. Cryptography provided a solid understanding of Transport Layer Security (TLS) when we leveraged Caddy for encryption. Cloud Security was important during the solution's migration to AWS and in applying the appropriate security restrictions. Machine Learning provided a foundation for understanding and evaluating various classification models. Natural Language Processing gave us a deeper understanding of transformers and their use as encoders for classification. Operating Systems provided understanding when deploying containers within a Linux environment. Web Application Security and Cyber 200: Beyond the Code informed the development of our policy framework and the security of prompts being sent over HTTP. Software Security gave us a proper understanding of how to secure the messaging application with input validation techniques and provided the foundation for our prompt injection detection capability.
One of the things Brandon and I felt strongly about was tying together what we learned across the MICS program into one cohesive solution. The final project was not based on a single class or technology, but on applying concepts from across the curriculum to build a working security platform.
Do you have any future plans for the project?
Sash: Yes, we plan to continue the project and have already begun brainstorming ideas for what comes next. The foundation of the system provides a great template for expansion. I am most excited about the opportunity to pilot the capability in a real environment and demonstrate that in-line AI governance can be effective, low-latency, and easy for users to adopt.
Brandon: We have a solid foundation coming out of the capstone, and I think our product fills a real gap in the AI governance space by enabling the safe, privacy-conscious use of AI. I’m interested to see what business audiences will think of our product.
How could this project make an impact?
How could this project make an impact?
Sash: I think this project can have a significant impact by providing security, visibility, and control over AI messages. AI adoption is growing exponentially, and new protocols are being developed that allow AI systems to communicate with other AI systems, devices, and users. As these interactions grow, organizations will need a way to secure their enterprise AI boundary. I believe an in-line security tool for AI traffic, similar to the role email security gateways play today, will become increasingly important.
Brandon: TrustLayer has the most impact for organizations in highly regulated or sensitive industries such as healthcare and defense, where teams want to use LLMs but the models are often not made available due to the risk of exposing sensitive data. Most companies are currently either blocking AI altogether or using a heavily sanitized version, which strips most of the utility. Our objective with TrustLayer is to provide balance by striking the right level of governance without hindering operations.
Now that you’ve finished your capstone, what advice would you give yourself at the start of the program? What do you wish Day One you knew?
Sash: I would tell myself to view each class as a building block rather than an individual course. Looking back, many of those concepts came together in our capstone project. I would tell Day One me to experiment more, keep building on ideas from previous classes, and think about how the different areas of cybersecurity can work together to solve a larger problem.
Brandon: There's a lot you can leverage from the classes throughout the program, and I think identifying your area of focus early and choosing courses with synergies adds a lot of value. For me, pairing the data science courses with cybersecurity was especially valuable when it came to building the project and more broadly in my day-to-day work.
