Zhen Zhou

AI/ML Engineer & Researcher
LLM Systems • Medical Imaging AI • Graph ML
[Email] [GitHub] [Google Scholar] [CV]

AI/ML engineer with 7+ years owning the full model development lifecycle, from dataset curation and architecture design through validation and HPC deployment, delivering production-ready PyTorch pipelines across large-scale biomedical datasets. Deeply passionate about large language models and generative AI: experienced in fine-tuning, prompt engineering, retrieval-augmented generation (RAG), and building agentic systems that bridge structured domain knowledge with foundation models. Practiced in version control, CI-compatible pipeline design, containerization, and SLURM-based distributed computing.

Hands-on builder of agentic LLM systems: developed a multi-agent orchestration pipeline for automated QC, harmonization, and report generation; built a graph-encoder-to-Llama-3-8B bridge for structured clinical report synthesis; and LoRA fine-tuned foundation models on domain-specific biomedical time series. Open-sourced BrainNetClass (80+ GitHub stars), a production-ready graph ML toolkit adopted by 15+ institutions globally. 32 publications (797 citations) in Nature Communications, PNAS, and NeuroImage.

Experience

AI Researcher/Engineer — MGH & Harvard Medical School, Boston, MA Aug 2023 – Present
  • Developed a generative U-Net pipeline integrating structural MRI and diffusion tractography to synthesize deformation fields for cortical surface alignment across 150K+ vertex manifolds; 5.5% improvement over baselines.
  • Built end-to-end LLM pipeline connecting DTI-derived brain connectivity with Llama-3-8B to synthesize structured clinical radiology reports via graph encoder and domain-specific prompting.
  • Built ReAct-style LLM agent orchestrating an 8-tool neuroimaging pipeline for automated QC, harmonization, and report generation.
  • LoRA fine-tuned Llama-3-8B on fMRI time series with multiscale decomposition and cross-attention reprogramming for brain connectivity extraction.
Postdoctoral Researcher — University of Pennsylvania, Philadelphia, PA Jan 2022 – Jul 2023
  • Developed a multi-site neuroimaging framework across 4,259 subjects; replaced statistical harmonization with a 3D CNN + attention model, reducing prediction error by 15.5%.
  • Contributed biomarker extraction pipelines to 4 high-impact publications linking MRI features to genetic architecture and genomic loci in Alzheimer's disease cohorts (Nature Communications, PNAS, EBioMedicine, Translational Psychiatry).
Research Assistant — University of North Carolina at Chapel Hill, NC Nov 2018 – Dec 2020
  • Designed and open-sourced BrainNetClass, a production-ready graph-based ML package for imaging biomarker extraction; independently deployed at 15+ institutions globally.
  • Built the first infant neuroimaging pipeline with automated QC and multi-layer temporal network analysis for the Baby Connectome Project.

Recent Highlights

  • 2026.01: Paper on deep-learning cortical registration guided by structural and diffusion MRI accepted at ISBI 2026 — invited for oral presentation.
  • 2025.05: Paper on harmonization of brain structural connectivity through distribution matching accepted at Human Brain Mapping.
  • 2024: Contributions to genetic architecture of multimodal brain age and genomic loci studies accepted at Nature Communications and PNAS.
  • 2024: MRI brain-age and Alzheimer's neuropathology study accepted at eBioMedicine.
  • 2023.02: Multiscale functional brain aging connectivity patterns accepted at NeuroImage.

Technical Skills

Languages PythonC/C++JavaMATLABShell
Deep Learning PyTorchTensorFlow/KerasScikit-learn
LLMs & Agents HuggingFacePEFT/LoRALlama 2/3BERTAutoGenLangChainRAGAgentic Systems
Infrastructure SLURM/HPCDockerAWSGitMLflowLinuxMLOps
Domain Graph MLHealth AINeuroimaging (MRI/fMRI/dMRI)Drug Discovery & GenomicsKnowledge Graphs

Education

Ph.D., Computer Science — Zhejiang University, China Sep 2014 – Dec 2021
B.Eng., Software Engineering — Northeastern University, China Sep 2010 – Jun 2014