cv

Safe RL, generative personalization, contrastive physiological modeling, scientific computing, uncertainty quantification, wearable health algorithms, selected publications, and deployed product work.

Basics

Name Zexin Liu
Label Medical AI and Wearable Health Algorithm Researcher
Email liu.zexin9518@gmail.com
Url https://zexinliu.github.io
Summary Advanced algorithm engineer at HONOR with a Ph.D. in Mathematics from the University of Utah. I combine scientific computing, uncertainty quantification, physiological signal modeling, safe reinforcement learning, generative modeling, and contrastive representation learning to build deployable medical and wearable health algorithms.

Work

  •
    Advanced Algorithm Engineer
    HONOR
    Research and development of medical and health-related algorithms for wearable devices, with emphasis on physiological signal modeling, representation learning, calibration, and product-level robustness.
    • Led commercialization of 24-hour continuous non-invasive blood pressure monitoring for the HONOR Watch 5 Pro.
    • Led commercialization of sudden cardiac arrest screening based on heart-rate deceleration capacity for the HONOR Watch 5 Ultra.
    • Worked on physiological signal modeling, multimodal representation learning, calibration, domain adaptation, and health-feature algorithm pipelines.
    • Developed applied project tracks in safe RL for physiological control, conditional generative personalization for cuffless BP, ECG/PPG contrastive learning for deceleration-capacity screening, and uncertainty-aware biomedical simulation.

Education

  • Beijing, China

    Bachelor's degree
    Beihang University
    Mathematics

  • Salt Lake City, Utah, United States

    Ph.D. in Mathematics
    University of Utah
    Mathematics; scientific computing and uncertainty quantification
    • Advisor: Professor Akil Narayan
    • Orthogonal polynomial recurrence algorithms for nonclassical measures
    • Multivariate Stieltjes algorithms and recurrence matrices
    • Polynomial chaos expansion and noninvasive uncertainty quantification
    • Open-source UQ software for computational biomedical simulations

Awards

Publications

Skills

Scientific Computing and Uncertainty Quantification
Orthogonal polynomials
Polynomial chaos
Gaussian quadrature
Biomedical simulations
UncertainSCI
Medical AI and Wearable Health Algorithms
Physiological signals
Photoplethysmography
Electrocardiogram
Blood pressure prediction
Deceleration capacity
Calibration
Domain adaptation
Generative Modeling and Representation Learning
Contrastive learning
Flow matching
Diffusion models
Multimodal learning
Signal denoising
Representation learning
Safe Reinforcement Learning and Physiological Control
Safe RL
Constrained MDP
Action shielding
Physiological control
Simulation-based evaluation
Treadmill control
Medical Large Models
Fine-tuning
Medical-domain adaptation
Healthcare evaluation
Knowledge-grounded modeling
Healthcare workflow modeling
AI Systems
Distributed training
Large language model inference
Paged KV cache
Tensor parallelism
vLLM

Languages

Chinese
Native speaker
English
Professional working proficiency

Interests

Research and Engineering Interests
Medical AI
Wearable health
Scientific computing
Uncertainty quantification
Safe reinforcement learning
Generative personalization
Multimodal contrastive learning
Large-model systems

Projects

  • 24H Cuffless Blood Pressure Monitoring
    Wearable PPG blood pressure monitoring project with personalization, calibration, generative modeling, and product-level robustness.
    • PPG blood pressure regression
    • Personalized calibration
    • Generative personalization
    • Calibration-aware modeling
    • Product deployment
  • Deceleration Capacity Screening from Wearable Signals
    Wearable heart-risk screening project based on deceleration capacity theory, ECG/PPG representation learning, and long-horizon physiological aggregation.
    • ECG/PPG representation learning
    • Deceleration capacity
    • Signal-quality-aware modeling
    • 24-hour aggregation
    • Wearable health-study deployment
  • SafeRunRL: Safe Reinforcement Learning for Adaptive Treadmill Control
    Simulation-based safe RL prototype for adaptive treadmill speed and incline control from wearable physiological signals.
    • Wearable state estimation
    • Personalized cardiovascular simulator
    • Constrained MDP
    • Action shield
    • Stress-test evaluation
  • Orthogonal Polynomials and Biomedical Uncertainty Quantification
    Ph.D. research on recurrence algorithms, polynomial chaos expansion, UncertainSCI, and reliable biomedical simulation uncertainty quantification.
    • Univariate recurrence coefficients
    • Multivariate Stieltjes algorithm
    • Polynomial chaos
    • UncertainSCI