Research

My work explores how to bring AI and cognitive-science insights into real-world education and health settings. At its core, I believe artificial intelligence should help people learn, not just automate tasks, by making the invisible processes of learning and decision-making visible, measurable, and actionable. I build systems that are clinically and pedagogically grounded, designed for real workflows, and evaluated with criteria that matter: reliability, interpretability, and impact on outcomes.

Current Projects

Adaptive Exercise Recommendation and Student Modeling (KUL-Rec)
In the Zitnik Lab, I develop systems that help learners find the next useful practice exercise as their understanding evolves. The Ken Utilization Layer (KUL-Rec) is a memory-augmented, neuroscience-inspired continual-learning method: drawing on the complementary learning systems theory from neuroscience, it maintains and updates an evolving memory of a learner’s needs rather than selecting exercises from static performance patterns alone. The goal is a recommendation that holds up under realistic classroom constraints—sparse, noisy interaction data and shifting learner understanding—and that stays interpretable and feasible within existing instructional workflows. The evaluation nuance matters here: in benchmarking, KUL-Rec outperformed baselines on retrieval-based metrics while scoring lower on AUC, a trade-off that is central to how the method should be positioned and assessed. The work is available as an arXiv preprint and is under review at the International Journal of Artificial Intelligence in Education.

Automated Short-Answer Grading and Feedback (ASAG)
Using large language models and natural language processing, I build tools that evaluate open-ended short-answer responses and generate feedback while keeping instructors in control. These systems are grounded in explicit rubrics and structured criteria, with an emphasis on grading reliability, rubric alignment, failure-mode analysis, and instructor oversight and override—so that automated assessment supports, rather than replaces, expert educational judgment. I led an end-to-end classroom deployment in the first-year Foundations course (2025) and a structured student focus group that surfaced clear, actionable needs, which in turn shaped follow-on work. This research was published in the New England Journal of Medicine: AI (2026).

Flipped-Classroom RAG Chatbot (VITAL)
I designed a retrieval-grounded conversational assistant (VITAL) that supports flipped-classroom learning—helping students prepare before class and consolidate afterward by asking questions and receiving responses grounded in their assigned materials. The design emphasizes citation discipline, evidence-grounded responses, and pedagogically useful explanations built for the constraints of real educational settings. VITAL was deployed in a graduate biochemistry course pilot (~35 students) and adapted for the TEECH (Technology Enabled Education for Community Health) continuing-education program, extending its reach to roughly 40 practising physicians serving remote, rural, and Indigenous communities.

Learning-Focused Syllabus Design
Faculty play a central role in shaping how AI enters the classroom. I built a syllabus evaluator that scores course syllabi against learning-science criteria and generates structured reports to support iterative, evidence-informed improvement. The project translates principles from learning science and curriculum design into measurable criteria, then turns those criteria into actionable feedback—lowering the barrier to rigorous course design while preserving instructor nuance, autonomy, and accountability.

Agentic Course Navigation with Domain Graphs (EduARD / King Project)
Building directly on what the Foundations focus group told us—students needed better ways to navigate dense course materials—I developed EduARD (Educational Adaptive Retrieval from Domain Graphs), an agentic AI framework for course-material support. Rather than functioning only as a chatbot that retrieves passages, EduARD organizes course content into a curriculum map or course-domain graph, answers student questions using that structure, and guides learners toward relevant materials and connections for deeper conceptual understanding. It is implemented on an Azure cloud platform with a PostgreSQL database backend to support scalable, structured retrieval. A classroom pilot has been advanced as a working prototype, with full deployment currently pending institutional platform and funding support, and the work continues as an active research direction. This line of work also includes a FAISS-based retrieval system over Foundations course materials and a GraphRAG baseline supported by a course-domain knowledge graph, used to study how structured, curriculum-aware retrieval can improve evidence-grounded educational responses.

Health and Imaging AI

Alongside my educational work, I continue research in biomedical imaging and clinical risk modeling. My doctoral research in Medical Biophysics at the University of Toronto focused on AI for breast MRI interpretation and risk prediction, integrating imaging and text data to capture subtle markers of disease and quantify longitudinal change. I have also worked on breast tissue segmentation, domain adaptation across imaging protocols, uncertainty-aware modeling, and language understanding in radiology reports—including BI-RADS-oriented pipelines for extracting structured signals from clinical text. These experiences continue to shape my approach to educational AI—reinforcing that transparency, robustness, and fairness are essential whether the data represent patients or students, and that evaluation must be tied to real decisions and real consequences.

Overarching Themes

1. Transparency and Explainability
   Whether in clinical research or education, tools must make sense to their users. I design models where reasoning and results can be examined, questioned, and trusted, with evaluation that prioritizes interpretability and reliability rather than performance alone.

2. Learner-Centered Design
   I focus on the cognitive journey of the learner: what they know, what they are ready to explore next, and how understanding changes through feedback, practice, and reflection. My goal is AI that supports learning as a process over time, not a single snapshot.

3. Scalable and Responsible Infrastructure
   I aim to build systems that scale thoughtfully and transparently. AI should amplify human expertise, not flatten it. The goal is infrastructure that supports growth, curiosity, and accountability—for students, educators, and clinical stakeholders alike.