Zhexian(Jako) Zhou

Zhexian(Jako) Zhou

Stay determined.

Zhexian Zhou (Jako) — Profile

I’m Zhexian (Jako) Zhou, an M.S. student in Artificial Intelligence Engineering (ECE) at Carnegie Mellon University, and a Research Assistant at CMU AirLab / Team Chiron. My work focuses on multimodal Vision-Language Models (VLMs) for embodied triage systems and informative 2D/3D skeletonization methods with strong geometric guarantees.

Contact

Downloads

Education

  • Carnegie Mellon University (CMU), Pittsburgh, PA
    M.S. in Artificial Intelligence Engineering (ECE), GPA: 4.0 — Aug 2024 – Dec 2025
    Lab: CMU AirLab / Team Chiron (Advisor: Prof. Sebastian Scherer)
    Coursework: Advanced Computer Vision, ML, Deep Learning, Trustworthy AI, Estimation/Detection/Learning, Systems & Tool Chains for AI
  • University of Electronic Science and Technology of China (UESTC), Chengdu, China
    B.E. in Software Engineering (Elite Program), GPA: 3.94 — Aug 2020 – Jul 2024

Research Interests

  • Multimodal VLMs (RGB-IR), VQA, embodied/edge inference systems
  • Geometric vision: distance transforms, medial axis/surface, topology-aware thinning
  • Hardware-aware ML: quantization, KV-cache optimization, efficient deployment

Research Experience

  • Research Assistant — CMU AirLab & Team Chiron (DARPA Triage Challenge) — Sep 2024 – Present
    • Built ROS2-based end-to-end edge inference + benchmarking pipeline for multimodal triage VQA (UAV/UGV), supporting SOTA VLM APIs (QwenVL, InternVL, Llama, VILA).
    • Proposed RGB-IR knowledge transfer and SFT fine-tuned NVILA, improving IR Hemo Acc: 36.6% → 54.1%.
    • Implemented quantization + multi-turn KV-cache optimizations on AGX Orin for real-time follow-ups (< 1s latency).
    • Developed Air-Infer for streaming ROS2 messages and vision tensors over gRPC/HTTP: github.com/JakoError/air-infer.
  • Researcher — 2D/3D Skeletonization (CMU AirLab) — Mar 2025 – Present
    • Proposed feature-aware envelope EDT with geometry-adaptive neighborhood interpolation, improving reduction rate 96.51% → 96.81% while increasing reconstruction IoU 88.14% → 99.59% (2D scikit-image blob dataset).
    Ours: skeletonization result
    Ours. Feature-aware envelope EDT + geometry-adaptive interpolation.
    MA (medial axis): skeletonization result
    MA (Medial Axis). Baseline method.
  • Undergrad Researcher — Semi-supervised Semantic Segmentation (UESTC) — Oct 2022 – Jun 2024
    • Outstanding Thesis (UESTC 2024): dynamic confidence + edge-aware semi-supervised learning for sparse segmentation; mIoU 70.5 → 76.0 → 83.9 → 86.9 as supervision rate 10 → 20 → 50 → 100.
    • Geometry-aware route-pattern recognition for tubular structures (e.g., vessels).
  • Undergrad Researcher — Memristor Computing (UESTC) — Jan 2023 – Jun 2024
    • Fixed-point FFT and evolutionary algorithms mapped to memristor crossbar circuits; validated via circuit-level simulations and noise-mitigation design.
  • Researcher — Tibetan Language Automatic Recognition (UESTC & TibetU) — Jun 2022 – Jun 2024
    • Built diffusion-based OCR/scene-text augmentation platform integrated into the project training pipeline.
    • MoST China 2030 Key R&D Program (No. 2022ZD0116100).
  • Research Intern — NeRF for Surrounding-View Reconstruction (Peking University) — May 2023 – Aug 2023
    • Adapted NeRF for urban scenes; trained on KITTI and transferred to nuScenes for improved 3D rendering.

Publications

  • Accepted (ICLR 2026): Luo, Y., Zhou, Z., Chen, H., Qiu, K., Savvides, M., Li, S., Wang, J.
    KnowledgeSmith: Uncovering Knowledge Updating in LLMs with Model Editing and Unlearning.
    arXiv: 2510.02392 · OpenReview: znnA2Opw6v
  • CAIT 2023: Z. Zhou et al. Design of Fixed-point FFT Algorithm Based on Memristor.
    DOI: 10.1109/CAIT59945.2023.10468999
  • ICCIP ’23 (Proc. 2024): X. Zhong, Y. Yu, C. Zhou, X. Wang, X. Feng, Z. Zhou, et al.
    A Matrix Coding Genetic Algorithm Based on Memristor for Image Edge Detection.
    DOI: 10.1145/3638884.3638895

Selected Projects

  • Air-Infer (ROS2 ↔ gRPC/HTTP streaming) — Robust transmission of ROS2 messages and vision tensors for inference pipelines
    github.com/JakoError/air-infer
  • MatSAR — C++ Math Toolkits — Cross-platform C++ toolkit (matrix core + ops + math functions); led development of 200+ algorithms (Linear Algebra / DSP / Calculus). (Open-sourcing in progress.)
  • C++ Database Management System — Server-client DBMS with core storage/query components
    github.com/JakoError/cppDBMS
  • Compiler Toolkits (SysY) — Flex/Bison compiler with IR generation and basic optimization
    github.com/JakoError/CompileStudy
  • DiaryInUESTC (Mobile App) — Diary/bookkeeping/memo app with AMap and MyBatis persistence
    github.com/JakoError/DiaryInUESTC

Skills

  • Languages: C/C++, Python, Java, MATLAB, SQL, JavaScript, HTML/CSS
  • Frameworks: ROS2, PyTorch, PyTorch Lightning, TensorFlow/Keras, NumPy, gRPC/REST, PySpark, Spring Boot
  • Tools: Git, Docker, AWS, VS Code, Visual Studio, GCC, CMake

Honors & Awards

  • Outstanding Undergraduate Thesis Award — UESTC (Jun 2024)
  • Outstanding Graduate Award — UESTC (Jun 2024)
  • Science & Technology Innovation Seedling Award — Sichuan Program (Oct 2023)
  • CASC Scholarship — China Aerospace Science and Technology Corporation (Dec 2022)
  • Academic Outstanding Student Scholarship — UESTC (Dec 2021, Dec 2022)
  • China Southwest Hackathon — 3rd Place (Final) (Oct 2020)

If you like, here is an old history → Old History