About
Hi! I’m Leo Sin (施延庆).
I am a visiting student at Westlake University, Intelligent Spatial Medicine Lab, supervised by Prof. Haoyang Li. I received my M.Sc. in Computer Engineering from The University of Hong Kong (supervised by Prof. Grantham Pang) and my B.S. in Electrical Engineering from Soochow University (supervised by Prof. Lu Bo).
My research focuses on 3D perception for surgical and biomedical scenes, spatial omics, and multimodal AI. I am particularly interested in applying generative AI and 3D reconstruction (NeRF/3DGS) to medical imaging and spatial transcriptomics.
Research Interests: Spatial Omics, 3D Reconstruction (NeRF/3DGS), Surgical Scene Analysis, Multimodal Medical AI, Embodied Intelligence, LLM Agent Systems
News
- 2026.02 Joined Westlake University as visiting student.
- 2026.01 Built Maze Mask at Global Game Jam 2026 — mask.colanns.me
- 2025.12 Started research internship at Tongji University CDI on robot questioning strategies.
- 2025.10 Working on autonomous driving digital twin with 3DGS + CARLA/Autoware.
- 2025.06 Joined Tab Next (HK) as algorithm engineer intern, developing medical AI chatbot.
Selected Projects
Spatial Omics — 3D Cell Atlas Reconstruction
Core research on 3D virtual slice interpolation for spatial transcriptomics. Synthesize virtual intermediate tissue slices between real 2D spatial transcriptomics sections to reconstruct 3D cell atlases. Working with Xenium sequencing data, H&E staining image to spatial omics mapping.
Medical RAG Chatbot — AI Agent for Imaging Center
Built ReAct agent-based medical chatbot for TST Imaging Center. Designed 7-tool architecture (business query, appointment, form analysis), CoT + FAISS retrieval, LoRA fine-tuning (SFT+DPO), vLLM acceleration (+70% performance), and Qwen2.5VL multimodal WhatsApp processing. CSAT on par with human agents, saving 20% labor.
AskThenRearrange — Robot Questioning Strategy via KCQ Encoding
Designed robot questioning strategy planning based on KCQ encoding from real human dialogues. Conducted 24-group 1v1 user experiments comparing result-oriented vs. process-oriented tasks. Found process-oriented tasks better suit user-preference strategies. Deployed on Unitree humanoid robot.
Paper: Task Matters: Investigating Human Questioning Behavior in Different Household Service for Learning by Asking Robots (in progress)
GitHub → Figma Slides →
Autonomous Driving Digital Twin — 3DGS + CARLA/Autoware
Built Image → COLMAP → 3D Gaussian Splatting reconstruction pipeline with multi-view preprocessing and camera pose recovery. gsplat-based training with PSNR/SSIM/LPIPS evaluation. CARLA simulation with vehicle, pedestrian, camera, LiDAR, GNSS, IMU deployment. Exploring digital twin for Autoware co-simulation.
Continuum Robot for Intestinal Surgery + SLAM Navigation
Built 6-stepper-motor continuum robot for intestinal tract with FTL precision testing via Raspberry Pi + STM32. A* path planning algorithm burned into STM32. Extended with LiDAR SLAM + ROS for autonomous scene reconstruction and obstacle avoidance. CAD-designed chassis platform.
Maze Mask — Global Game Jam 2026
Multiplayer online pixel-art game: multiple thieves navigate pathways, collect coins, find exits and escape the maze. Masks grant different abilities with roguelike power-up elements. Responsible for map design, planning, and mask feature implementation.
Play Online → Bilibili →
UAT-NET — Multimodal Blood Pressure Estimation
Attention-based mechanism for blood pressure estimation from multimodal PPG and ECG signals. Reproduced and improved upon existing methods for reconstructing BP signals from physiological waveforms. Evaluated with MAE/RMSE on PulseDB dataset using 3× NVIDIA RTX 4090.