Security and Isolation of AI Agents
We build AI tools and assistants for streamlining learning and academic tasks for University students, faculty, and staff. We put emphasis on ensuring these tools are privacy-conscious, by employing our lab's research when appropriate as well as other best practices.
BU undergraduates can reach out to Kinan if they are interested in this effort. They can also look into BU Spark and CDS's practicum courses ( DS519, DS539, and DS549) for more hands-on software and data engineering and development opportunities
Active Projects
Our DS593: Privacy-Conscious Computer Systems course is a reading intensive course, where students read recent technical research papers from leading conferences in this field. Unfortunately, students increasingly rely on LLMs in their coursework, including readings, often to their own determent: in addition to obvious academic integrity issues, the LLM may produce incorrect or shallow outputs, and students may utilize it to skip doing important work.
We are building a responsible paper-reading assistant, called PaperBuddy, that can act as a thinking buddy while students are reading, rather than a substitute for it. We carefully design PaperBuddy's prompts to assist students in their thinking, rather than giving them direct answers or summaries. PaperBuddy logs student interactions with LLM, to allow course staff to review and correct any inaccurate answers. PaperBuddy reports common issues, difficulties, or questions the students have about the paper to the teaching staff.
Participation in PaperBuddy and in related experiments to measure its efficacy is optional for students. We plan to use our K9db and Sesame in PaperBuddy, to ensure that it complies with students access requests and correctly abides by their privacy and consent preferences. We will also use PaperBuddy as a test bed for some of our on going research, including Tahini and LLMMarshal.
We are building an LLM-powered file analysis platform and pipeline that BU students, faculty, and staff can use. This platform handles different types of files, such as academic syllabi, resumes, and others, and allows integrating LLM-powered analysis within larger workflows. We are building on top of BU's TerrierGPT to ensure that user data remains private, does not get stored by third parties, and does not get used for training. We are interested in expanding this effort, and collaborating with others at BU, to build a suite of privacy-conscious productivity tools (a kind of "TerrierGPT studio").