Research

Research Statement

I build resource-efficient multimodal machine learning systems that turn noisy real-world sensor data into reliable health and behavior insights. My work sits at the intersection of machine learning, unobtrusive sensing, mobile health, and edge deployment, with a focus on systems that can operate outside laboratory settings under constraints on compute, power, privacy, and user burden.

The core question driving my research is: how can intelligent sensing systems make accurate, timely decisions on-device when data are sparse, multimodal, and highly variable across people and environments? I approach this question by designing end-to-end pipelines that connect sensing hardware, signal processing, model design, and deployment-aware evaluation.

My research contributes across three connected areas:

• On-device multimodal integration: I develop adaptive sensing pipelines that integrate thermal, RGB/depth, IMU, and physiological signals while selectively activating higher-cost sensors only when needed.
• Efficient model design and deployment: I design lightweight deep learning models and use neural architecture search, conditional computation, and deployment-aware optimization to support real-time inference on resource-constrained devices.
• Robust real-world health modeling: I build and evaluate models that generalize across users, environments, and longitudinal deployments, with an emphasis on interpretability, reliability, and clinically meaningful outcomes.

Across projects in dietary monitoring, eating behavior understanding, energy expenditure estimation, stress sensing, and smoking topography estimation, I aim to bridge algorithmic innovation with deployable systems. I am especially interested in research roles where I can develop practical ML systems for health, wearables, mobile sensing, and edge AI, moving ideas from prototype models to robust real-world products while grounding system design in users’ needs for reliable and truly useful tools.