Research

I designed and deployed MerryQuery, an educational assistant that grounds its responses in course materials, cites sources, structures explanations as steps, and gives instructors control over how much of a solution it reveals. It has supported more than 750 students across three semesters and received the AAAI 2025 Best Technical Demonstration Award.

In a 14-week classroom study, 162 consenting students generated more than 4,000 queries. Seventy-one percent reported that the assistant helped them learn, yet students ranked teaching assistants and course materials above it. These findings motivate closer study of the instructional choices behind AI help.

I developed HelpLens, an eight-dimension framework and annotated dataset comparing teaching-staff and LLM responses to 172 authentic programming questions. The model offered code more often; staff more often calibrated help to the type of question. This work characterizes assistance and informs future studies of its effects on learning.

In interviews with 16 early-adopter software developers, we found that LLMs supported implementation, refactoring, and debugging, while people continued to check, revise, and often discard the output. Human judgment and foundational knowledge remained central to their work.

I led a systematic review of 52 studies of student-facing LLM tools in computing education. Most evaluated students' perceptions; fewer measured academic performance. None cast AI as a teammate, opening questions about how AI might support peer learning.

I lead ExLAIM, an NSF ExLENT-funded co-design program. In Summer 2026, nine K–12 teachers and seven undergraduate developers built five classroom tools on MerryQuery. Teachers asked for human checkpoints: editable drafts, review before release, and control over when AI contributes.

I evaluated programming struggle and progress detectors against think-aloud protocols, screen recordings, and learners' reports. The detectors sometimes mistook copied starter blocks for progress or unusual correct strategies for struggle. Logs show behavior, but they do not fully explain a learner's reasoning.

In a quasi-experiment with 199 sixth-graders, students using Parsons problems (scrambled code blocks to reorder) were scored higher on the science task, spent less time coding, ran their programs more often, and stayed closer to a correct solution. The scaffold reduced the search for blocks and left room to test and explore.

I also contributed to Jigsaw, a visual tool for decomposing and planning programming problems, recognized with the IEEE VL/HCC 2024 Best Research Paper Award.