xinhezhang [at] g.harvard.edu · Generative & Multimodal AI
I am an AI researcher and PhD candidate in Electrical and Computer Engineering at Harvard University, based at Massachusetts General Hospital, where I work with Prof. Na Li and Prof. Quanzheng Li.
Before Harvard, I received BS and MS degrees in Electrical and Computer Engineering from Carnegie Mellon University, and worked as a software engineer at Google, Meta, Duolingo, and JP Morgan Chase.
Seeking AI Researcher / Postdoctoral roles starting January 2027.
I focus on generative modeling and multimodal representation learning, with experience in efficient generative architectures, large-scale foundation-model adaptation, and agentic AI systems.
Our Science paper on cyborg pancreatic islets was featured in the press — showing how flexible bioelectronics and spatial transcriptomics together reveal the maturation of transplanted human islet cells.
Our Science paper on flexible nanoelectronics for transplanted cardiomyocytes was covered by Medical Xpress — combining bioelectronics, machine learning, and spatial transcriptomics to study how transplanted heart cells give rise to arrhythmias.
Third Place for Poster Presentation at the Eric and Wendy Schmidt Center Biomedical Science & AI Symposium, for a closed-loop framework pairing Gaussian Process Bayesian optimization with flexible bioelectronics to optimize stimulation protocols for cardiac organoid maturation.
Early-Career Scholar Honoree and Best Poster Award Runner-Up at the NIH BRAIN NeuroAI Workshop, for a hierarchical generative model that jointly captures stimulus-evoked neural dynamics and long-term representational drift in large-scale single-neuron recordings.
arXiv · 2026
Efficient 3D medical image synthesis by learning inter-slice feature trajectories from 2D generators, at roughly 135× lower inference cost than comparable 3D methods. Evaluated on BraTS 2023 and SynthRAD2023 across unconditional generation, missing-modality MR synthesis, and MR-to-CT translation.
bioRxiv · 2024
A closed-loop platform integrating flexible bioelectronics with Gaussian Process Bayesian optimization to adaptively control electrical stimulation of cardiac organoids, reaching ~70% higher conduction velocity than fixed-parameter baselines.
Science · 2026
Tissue-like stretchable electronics implanted during organogenesis of stem cell-derived pancreatic islets enable months-long, single-cell-resolved electrophysiology, revealing two major electrical states distinguished by the glucose threshold for action potential firing.
Science · 2025
Ultra-flexible nanoelectronic meshes implanted alongside transplanted human iPSC-derived cardiomyocytes resolve arrhythmogenic dynamics at single-cell resolution over weeks in vivo, and show how coinjection with RADA16 promotes sarcomere organization.
IEEE ITSC · 2021
Uses the CARLA simulator to generate training data for rare road object detection, improving sim-to-real transfer for autonomous driving perception.
arXiv · 2026
A lightweight 2D-to-3D generative framework that learns inter-slice feature trajectories to produce coherent high-resolution 3D medical images at much lower inference cost than fully volumetric models.
PLOS Computational Biology · 2026
A systematic benchmark of spike source localization algorithms on simulated and experimental ground-truth datasets, identifying tradeoffs in accuracy, robustness, and runtime for long-term neural recordings.
Science · 2026
Implanted stretchable electronics longitudinally track single-cell electrical activity in human pancreatic organoids, revealing maturation trajectories of stem-cell-derived α and β cells.
bioRxiv · 2026
Growth-adaptive spring electronics maintain stable neural interfaces during brain growth, enabling weeks-long tracking of the same neurons across postnatal development.
IEEE CDC · 2025
Adapts instrumental-variable and bias-compensation methods to learn linear dynamical systems from noisy observations with improved finite-sample guarantees.
Science · 2025
Flexible nanoelectronics monitor transplanted human iPSC-derived cardiomyocytes and identify the electrophysiological mechanisms underlying graft-induced arrhythmogenesis.
Advanced Intelligent Systems · 2025
Riemannian geometry on covariance features classifies intracortical brain states efficiently and interpretably, outperforming conventional baselines at much lower training cost.
bioRxiv · 2025
BehaveAgent, a multimodal AI agent that automates behavior analysis from video across species and tasks without task-specific retraining or manual annotation.
bioRxiv · 2025
A plastic–elastomer heterostructure platform combining mechanical robustness with tissue-level flexibility for scalable long-term brain-computer interfaces.
bioRxiv · 2025
STAgent, an autonomous multimodal agent that automates spatial transcriptomics analysis and charts in vivo maturation of hPSC-derived pancreatic tissues.
bioRxiv · 2025
An autonomous AI agent for spike sorting that streamlines preprocessing, detection, clustering, and validation to improve scalability and reproducibility of neural data analysis.
bioRxiv · 2025
A multimodal AI agent that integrates LLMs, VLMs, and domain tools to automate flexible bioelectronics design, fabrication planning, defect inspection, and signal analysis.
bioRxiv · 2025
Pretrained vision-language models link extracellular electrophysiological features with molecular cell identities, enabling AI-assisted cell-type-specific brain-computer interfaces.
bioRxiv · 2024
A closed-loop AI-cyborg platform combining flexible bioelectronics with Bayesian optimization to adaptively modulate organoid maturation in real time.
IEEE IEDM · 2024
Soft, tissue-level flexible bioelectronics for minimally invasive in vivo neural recording and stimulation in the brain and spinal cord.
arXiv · 2024
DriftNet, a lifelong-learning framework inspired by representational drift that mitigates catastrophic forgetting across image, language, and large-model adaptation tasks.
bioRxiv · 2024
Flexible brain-machine interfaces are used to study and potentially realign representational drift in mouse visual cortex over longitudinal neural recordings.
Nature Communications · 2023
UnitedNet, an explainable multi-task deep learning model for integrating single-cell multimodal data and performing cross-modal biological inference.
IEEE ITSC · 2021
Synthetic data generated in the CARLA simulator improves rare-object detection when real labeled examples are scarce.
No publications in this category.
† equal contribution · ‡ corresponding author