Research portfolio

Projects that connect AI, sensing, and human behavior.

My work spans virtual sensor generation, multimodal learning, wearable computing, and inclusive human-centered systems. These projects share one goal: making behavioral AI more scalable, realistic, and useful.

IMUdrape: a generated jogging motion with body-mounted virtual IMUs (blue) and garment-mounted virtual IMUs on a physics-simulated sensor jacket and trousers (orange), next to the matched MIG and LIP garments in front and back views
Smart garmentsUnder review

IMUdrape

Virtual IMU data for sensors sewn into clothing.

Sensors in clothing move with the fabric, not the body. IMUdrape generates motion from text, dresses the body in a simulated copy of the deployment's garment, and reads virtual IMUs off the fabric at the real sensor positions. On four clothing-sensor datasets it improves activity recognition and pose estimation over real data alone and over body-mounted virtual data.

4 real datasets7–46% macro-F1 gain with 1 real subjectPaired garment and body signals
Project page Paper Code
EduGage study setup: a participant wearing a Muse S Athena EEG headband, eSense earbuds, Microsoft Band 2, tau-Ring, and Polar H10 chest strap, seated at the video-learning workstation
Engagement sensingIMWUT 2026

EduGage

Can sensing help us understand when a learner is truly engaged?

EduGage combines PPG, ECG, EDA, EEG, IMU, heart rate, temperature, eye tracking, and interaction data with in-situ engagement reports. The project contributes both a multimodal dataset and practical evidence for lightweight sensing combinations.

16 participants9+ signal types0.80 MAE83.18% within-1 accuracy
Project page Paper Code Dataset
AgentSense: an LLM-driven agent watching TV on the sofa in a simulated VirtualHome living room
Embodied AIAAAI 2026

AgentSense

Smart-home sensor data, generated by agents that live there.

LLMs create diverse personas and routines; embodied agents execute them in an extended VirtualHome simulator instrumented with ambient sensors. Pretraining on this data improves recognition, especially when real data is limited.

5 real datasetsPrivacy by design
Project page Paper Code
IMUGPT: LLM-written activity descriptions become generated 3D motion and virtual forearm accelerometer signals
Cross-modality generationISWC 2023 · IMWUT 2024

IMUGPT

Describe an activity; generate wearable motion data.

An LLM writes diverse descriptions of an activity, text-to-motion models turn them into 3D motion, and the motion is converted into virtual on-body IMU signals for training HAR models. IMUGPT 2.0 adds a motion filter for implausible clips and a diversity metric that tells you when to stop generating.

1.0 ISWC 2023 · Best Paper Honorable Mention2.0 IMWUT 2024
Project page Paper (1.0) Paper (2.0) Code
StitchSense: sensing from a smart garment is used as the source domain and transferred to a smartwatch for activity recognition
Smart garmentsISWC 2026

StitchSense

Learn from the garment, recognize on the watch.

Uses sensing from a smart garment as a source domain for watch-based human activity recognition, so data from clothing can help the device people already wear.

Paper
Wearable IMU emotion recognition system diagram
Affective computingHASCA 2024

Emotion recognition on the go

Personalized emotion recognition using wearable IMUs, with virtual motion data supporting model pretraining and adaptation.

Paper
Smart rings used for fingerspelling recognition
Accessible interactionASSETS 2023

FingerSpeller

Camera-free American Sign Language fingerspelling recognition and text entry using instrumented smart rings.

Paper Data
Wheelchair transportation mode detection illustration
Inclusive sensingISWC 2024

Wheelchair transportation sensing

Comparing data-collection strategies for transportation-mode detection with wheelchair users, recognized with a Best Paper nomination.

Paper