BionicFace is an assistive wearable research project for people with chronic facial paralysis caused by nerve damage. We are exploring glasses-based facial sensing and closed-loop electrical muscle stimulation to support real-time upper-face symmetry during social expression and communication, with a focus on hardware integration, calibration, comfort, safety, and everyday use.
OpenRing Companion is an independent Android companion app for Tsinghua’s Tau-Ring smart ring. It replaces a data-collection-oriented interface with a local-first health dashboard that turns passive ring measurements into interpretable daily scores, trends, and intraday views while keeping health data and exports on the user’s device.
The Information Transfer Rate Atlas is an independent comparison site for communication and control interfaces across HCI and BCI. We recalculate information transfer rates from original papers and system figures, providing step-by-step derivations and conservative cross-modal comparisons with tighter bounds for language and pointing tasks.
Wardian is a local-first desktop habitat where work done with CLI agents becomes visible, durable, and malleable. It preserves real provider sessions, project context, coordination, and evidence, then lets repeated work become reusable prompts, skills, agent classes, and automations.
This ongoing project develops an adjustable vibrotactile wrap and a planned two-phase study for lower-limb spasticity after stroke. The study will compare stimulation locations using electrophysiological and functional measures, then compare stimulation at rest with stimulation during gait training; no outcomes are claimed before data collection is complete.
EchoForce is a wristband that estimates grip force from acoustic reflections produced by subtle skin deformation above the forearm flexor muscles. We evaluated both user-dependent and user-independent models across sessions and hand orientations.
AppetiteCheck examines momentary, non-invasive vagus nerve stimulation as a low-attention intervention during eating. In a controlled study, participants ate less and more slowly under stimulation than under sham while reporting similar post-snacking satiety.
BreathePulse is a laptop-mounted breathing guide that uses pulsed airflow as an implicit cue. In a controlled N-back study, airflow guided slower breathing most consistently during the easier task, while adding workload and producing little measurable change in stress.
Touché explores flexible, low-cost wearable electronics for delivering vibrotactile stimulation to people with post-stroke upper-limb spasticity. The project focuses on making extended stimulation easier to use at home and during daily activity.
As part of the BCI Society’s Postdoc & Student Committee, I helped develop tutorials for the BCI Meeting and began organizing the material in a reusable repository. The initial machine-learning notebook exposes the choices behind preprocessing, features, models, and evaluation rather than presenting only a final classifier score.
At Microsoft Research, we collected 48,000 trials from 16 participants across eight closed-mouth tongue gestures using sensors in commercial head-worn devices. A multimodal model reached 94% accuracy, and a follow-on interaction technique used tongue gestures to confirm gaze targets without dwell for subtle, hands-free selection.
In a survey experiment with 51 Georgia Tech undergraduates, we tested whether the stated source and platform of a COVID-19 public-health message affected its credibility. Messages attributed to public-health agencies were rated above those with an unknown source; the text-only platform manipulation produced no significant effect.
This coursework prototype combined mobile controls, burner reminders, an adjustable cooking surface, and a side-loading oven to explore support for memory, reach, and posture in the kitchen. A four-person Wizard-of-Oz study identified useful interactions and substantial usability issues, but did not include the intended older-adult population.
Through roughly 50 hours of participant observation and interviews with six community members, we studied how people learned, collaborated, formed roles, and handled conflict in Horizon Worlds. Its creator community operated as a community of practice, while moderation and the ephemerality of encounters shaped whether its social spaces could remain dependable.
In a Neuromatch Academy collaboration, we analyzed Human Connectome Project fMRI data from 100 participants across 360 cortical regions. Contrasting 2-back with 0-back tasks isolated increased frontoparietal activity and reduced activity in parts of visual cortex, but did not establish the direction of information flow between regions.
We developed passive haptic rehearsal: a wearable glove, light-up keyboard, and learning portal that supplement active piano practice with tactile repetition between sessions. In the primary analysis of 20 participants with complete, valid data from a 36-person in-home study, passive rehearsal increased the modeled learning rate for note accuracy by 49.7%, but did not significantly improve rhythm.
BrainBraille is an attempted-movement brain-computer interface that maps combinations of body regions to language. For my undergraduate thesis, I explored a wearable fNIRS implementation and examined whether transitions between movements could expand the command set while improving classification accuracy.
We investigated whether the structure of warehouse order-picking tasks could provide weak supervision for object recognition from egocentric video. The system learned visual object models from repeated picks without frame-level labels and was evaluated on independent orders.
In this computer-vision course project, we built an early-fusion convolutional model for estimating 21 hand keypoints from synchronized stereo images. The model did not converge to a usable estimator; a persistent coordinate offset made the project a practical lesson in validating camera geometry independently of end-to-end training loss.
SilentSpeller is a hands-free text-entry system that senses tongue contact through capacitive electrodes in a dental retainer. We evaluated silent spelling with a 1,164-word vocabulary while participants were seated and walking, as well as in live text entry.
In this machine-learning course project, we represented Kepler K2 light curves using generic time-series features and compared five classifiers on a curated set of 685 stars. A random forest produced the strongest result, with an F1 score of 0.83 and cross-validated accuracy of 0.80 ± 0.06; the model ranked candidates rather than confirming new planets.
For RoboJackets’ autonomous ground vehicles, I developed motor-control and path-planning software and built a full-scale Gazebo reproduction of the competition course. I later managed the mechanical, electrical, and software subteams as the project integrated perception, localization, planning, and recovery behavior into a single outdoor robot.
I designed a recurrent reinforcement-learning and simulation framework for an autonomous robot in the 2018 DJI RoboMaster AI Challenge. Team RCMakers qualified as a finalist and placed 11th at ICRA 2018.
For the 2017 IEEE humanitarian-demining robotics challenge, our high-school team combined a low-cost mobile robot, dual-coil metal detector, decision-tree classifier, and ROS navigation stack. I worked on mine detection and supporting research; the team placed third and demonstrated the system at ICRA 2017.
