Sense · Infer user and environmental states
Wearable Multimodal Sensing
Building compact multimodal wearables with sensing modalities such as Electrical Impedance Tomography (EIT) and IMUs to infer user and environmental states.
About
Embodied AI through Wearable Interactive Systems
I am a second-year Ph.D. student in Computer Science at Northwestern University (Epistemic Decisions Lab), advised by Prof. Jessica Hullman. My earlier projects built wearable sensing and stimulation; at Epistemic Decisions Lab I design context-aware, proactive and mixed-initiative assistants for human-AI joint decision making.
My papers have appeared at ACM CHI, UIST, CVPR, and NeurIPS. I have given invited talks at MIT, UIUC, UMichigan, Columbia, and UChicago, and invited technical demos at Qualcomm and Google.
Previously, I was a research assistant with Prof. Pedro Lopes at the University of Chicago (HC-Integration Lab), with Prof. Karan Ahuja at Northwestern (SPICE Lab), and with Prof. Teng Han at ISCAS. I earned an M.S.E. in Robotics at the University of Pennsylvania (GRASP Lab).
News
- Nov 2026 Gave a lightning talk at the HCI Open House (Post-UIST 2026) at the University of Michigan.
- Nov 2026 Presented EITWatch at UIST 2026 in Detroit.
- Oct 2026 Gave an invited talk on embodied AI in wearables at the Computer-Enabled Abilities Lab at Columbia.
- Oct 2026 Our peer-review checklists paper received a Best Paper Honorable Mention at the NeurIPS 2026 Workshop on AI-Native Academia.
- Oct 2026 Gave an invited talk on multimodal sensing for the hand at the HCI Seminar at UIUC.
Research
My research explores Embodied AI in Wearable Form Factors, building context-aware assistants that act proactively or share initiative with the user.
- Build multimodal wearables that sense the wearer and their surroundings.
- Formalize when and how to take initiative as a decision-making problem, with models of human behavior as priors.
- Design on-body haptics that translate assistance decisions into physical interventions.
Decide · When and how to assist
Mixed-Initiative Assistance
Formalizing when and how an assistant should share initiative with the user as a decision-making problem, with models of human behavior as priors, to support people under uncertainty.
Intervene · Deliver physical assistance
On-Body Haptics
Designing compact on-body interfaces that deliver assistance through haptic feedback to support human perception and action.
Publications
* denotes equal contribution
Reviewer lottery error arises when a small set of peer reviews raises only some of the issues that a larger pool of eligible reviewers would. We show this error can be reduced by surfacing a checklist of criticisms learned from a historical review corpus. Our LLM-assisted pipeline generates a tree of checklists at different granularities, then selects the one that maximizes a theoretically-motivated measure of correlated agreement among reviewers. In simulated review with LLM reviewers on NeurIPS 2021 and ICLR 2022 data, the learned checklist reduces the mean squared error (MSE) of a three-reviewer committee by 58% and 21%, respectively. In a 20-participant human study, the learned NeurIPS checklist lowers the MSE of a paper ranking against the NeurIPS 2021 order from 0.48 to 0.22.
Wrist Electrical Impedance Tomography (EIT) can sense hand gestures from muscle- and tendon-driven impedance changes without cameras, but prior wrist-EIT systems wrap electrodes beyond the watch-back contact patch and rely on separate analog front ends. We present EITWatch, the first wrist-EIT system built around smartwatch case-back geometry. Eight planar electrodes in a compact ring read impedance from the skin patch the watch already occupies; because a planar array cannot encircle the wrist, EITWatch uses multi-depth scanning to probe tissue along multiple current paths. A user study shows that this watch-back form factor can support both macro- and micro-gesture recognition, pointing toward unobtrusive EIT input on everyday wearables.
Inertial odometry using only IMUs offers a lightweight approach to human motion tracking for augmented reality and wearables, but learning-based methods often drift because they lack explicit models of human motion dynamics. We propose MARIO, which grounds inertial odometry in human kinematics through a learned pose prior inferred from IMU data, promoting physically consistent motion during state propagation. We further develop a sensor-fusion framework that incorporates auxiliary signals from lightweight sensors already available on commercial AR glasses, including magnetometers, barometers, and secondary IMUs. Evaluated on a large-scale motion dataset, MARIO substantially reduces positional drift compared with strong IO baselines and improves robustness across diverse motion conditions, pointing toward lightweight, camera-free human tracking that unifies kinematic priors with multimodal sensing.
Sensory substitution devices enable users to perceive visual information through other modalities, such as touch. However, most existing devices place the camera at the user's eyes (head-mounted), which limits the ability to coordinate manual interactions. We propose "seeing and feeling" from the hand's perspective to enhance the flexibility and expressivity of sensory substitution. To this end, we engineered a back-of-the-hand electrotactile display that renders tactile images from a wrist-mounted camera, allowing users to feel objects while reaching and hovering. In a user study with sighted and blind or low-vision participants, we compared our hand-centered perspective against traditional head-mounted views in manipulation tasks (e.g., handling bottles, soldering). Results indicate that while both perspectives yield comparable performance, participants preferred the flexibility of the hand's perspective, which supported more ergonomic object manipulation strategies.
Electrotactile stimulation offers a promising path for high-resolution wearable haptics but has traditionally struggled with integration into soft, everyday textiles. We present TacTex, a textile interface that seamlessly integrates high-density haptic feedback and touch sensing. By employing a novel multi-layer woven structure that separates conductive electrodes with non-conductive yarns, TacTex achieves a sensing and actuation resolution of 512 × 512 with electrode spacing as small as 2mm. The system includes a custom driving board that enables precise spatial and temporal control of electrical stimuli while simultaneously monitoring voltage changes for touch tracking. Our technical evaluation and user studies demonstrate that TacTex can render a wide range of haptic effects, including complex static and dynamic patterns, effectively bringing high-fidelity haptic feedback to everyday clothing.
Teaching & Service
GradMAPS Mentor
Peer mentor in The Graduate School's GradMAPS program, supporting incoming first-year Ph.D. students as they acclimate to Northwestern.
Reviewer
Serving as a reviewer for CHI (human-computer interaction), IUI (intelligent user interfaces), and IMWUT (ubiquitous computing).
Teaching Assistant: ESE5190 Smart Devices
Graduate course at University of Pennsylvania covering embedded systems, IoT, and smart device development.
Code Instructor
Teaching Python & Arduino programming fundamentals to middle school students at Fife-Penn CS Academy.
More
Fun projects
-
-
-
-
Fun facts
- Blogger HCI and academic life, more than 5,000 followers.
- Saxophone player Paid player. More than 10 years of practice.
- Truck driver Drove from Chicago to NYC overnight without resting.
- Professional Gamer Fruit Ninja top 20 worldwide; NBA 2K top 500 nationwide.
Contact
