Apr. 2026 — Present
Internship · Lenovo
Inference acceleration for on-device LLMs
Profiling and optimizing local model inference with the LLM group.
Hello, I’m Zhiyi.
I build visual learning systems that connect research ideas with real-world imaging.
I am a second-year M.S. student in Computer Science and Technology at ShanghaiTech University. My current research focuses on sparse-view 3D ultrasound reconstruction with Prof. Rui Zheng. I am also an intern at Lenovo, working on inference acceleration for on-device large language models.
Previously, I worked with Prof. Sibei Yang at SooLab on multimodal learning, zero-shot human-object interaction detection, and view-invariant video understanding. My broader interests include computer vision, computer graphics, generative models, and efficient AI systems.
Experience
Research and industry
Apr. 2025 — Present
Research Assistant · BMU
Ultrasound image understanding · Prof. Rui Zheng
Working on pose-denoised freehand 3D ultrasound and sparse-view 3D reconstruction.
Aug. 2023 — Apr. 2025
Research Assistant · SooLab
Visual understanding · Prof. Sibei Yang
Explored multimodal LLMs, zero-shot HOI detection, and view-invariant video understanding.
Selected work
Research and projects
Sparse-view 3D Ultrasound Reconstruction
Reconstructing dense 3D ultrasound volumes from sparse 2D scanning trajectories for faster, motion-robust acquisition.
- Built a conditional generation pipeline with a 3D VQ-VAE and a 3D rectified flow model.
- Achieved strong results on the TUS-REC2024 dataset.
- Currently integrating IMU measurements toward an affordable 3D ultrasound system.
Zero-shot Human-Object Interaction Detection
Locating and recognizing interactions as <human, action, object> triplets, including combinations that were not seen during training.
- Designed a CLIP adapter and verb-focused feature extractor to improve action recognition.
- Introduced human-object interaction prototypes to address long-tail categories.
- Reached state-of-the-art results on unseen-verb, unseen-object, and selected unseen-composition settings.
Realistic Smoke Rendering
A renderer supporting both surface and participating-media rendering, developed for the Computer Graphics course at ShanghaiTech.
- Implemented Monte Carlo path tracing for physically based surface rendering.
- Used delta tracking for medium sampling and ratio tracking for transmittance estimation.
- Integrated volume and surface transport within a unified rendering pipeline.
Earlier work
Deep learning coursework
Nov. 2022 — Jan. 2023
Pixmath
An image-to-LaTeX formula recognizer using a CNN-GRU pipeline with attention, developed for a deep learning course project.
Background
Education and tools
Education
2024 — 2027 expected
M.S. in Computer Science and Technology
ShanghaiTech University
2023
Exchange Student · LAS
University of Illinois Urbana-Champaign · GPA 3.89/4.0
2020 — 2024
B.Eng. in Computer Science and Technology
ShanghaiTech University · GPA 3.55/4.0
Skills
- Languages
- Python, C, C++
- Tools
- PyTorch, Git, Linux, Bash, LaTeX
- Visual computing
- Image registration, depth estimation, generative modeling, rasterization, path tracing
- Languages spoken
- Mandarin (native), English
A question worth exploring
Can consciousness be created?
Engineering
Science
Philosophy
I am an adventurous and curious researcher who enjoys exploring beyond familiar boundaries. My work spans computer vision, computer graphics, multimodal learning, and medical imaging, including 3D ultrasound reconstruction. I am especially interested in connecting ideas across fields and turning unfamiliar questions into working systems.
I've also experienced basic training in philosophy and linguistics, which led me to think more deeply about intelligence, consciousness, meaning, and the relationship between semantics and pragmatics. These experiences convinced me that artificial intelligence is not only an engineering challenge, but also a question about language, knowledge, and the nature of mind.
My long-term vision is to explore whether conscious intelligence can truly be created—or whether the question itself may ultimately be unspeakable. I hope to work across engineering, science, and philosophy, building intelligent systems while continuing to ask what, exactly, we are creating.