Imperial College London
MSc Computing (Artificial Intelligence and Machine Learning)
Yuzhe Shi
Software Engineer and Amateur Computer Scientist
Imperial College London · MSc Computing (Artificial Intelligence and Machine Learning)
me@shiyuzhe.comGitHub / bkmashiroyuzhes.comLinkedIn / dylan030About
My current research focuses on AI infrastructure, particularly efficient execution environments for AI workloads, and unsupervised pretraining for world models. The first asks how model-generated tasks and tool calls can run safely, concurrently, and with low overhead; the second asks how environment structure can be learned from interaction data without task labels or rewards. Diffusion watermarking and faster diffusion sampling were earlier research projects.
I like cooking, especially baking and desserts. I built Zao because I did not want to touch my phone with oily hands while cooking. I keep the recipes I have made at wok.yuzhes.com.
I also play a lot of Minecraft and maintain RedScript and MapForge. The rest of my work is in the project archive.
I have a British Shorthair called . I am learning Japanese, which was also part of the original idea for Kotodama.
Education
South East Technological University
Bachelor of Science (Honours) in Software System Practice
First-Class Honours
Nanjing University of Information Science and Technology
Bachelor of Engineering in Software Engineering
GPA 89.9/100 · Ranked 2nd in the department (2/80)
Companies and teams
Shisho Guard (四象神户科技有限公司)
Co-founder / CTO · Nanjing
Co-founded a blockchain security company and led product and full-stack development for its audit system. The system used graph neural networks to infer blockchain address identities; the method was later granted a Chinese invention patent.
Developer / Administrator
Maintained the university judging system used for ACM/ICPC training, contests, and coursework. Worked on the judging backend, resource isolation, and scalability.
NUIST ACM/ICPC team
Team member
Trained in algorithms and competed in programming contests.
Selected projects
View all 13 projects →Connects NPC memory, world state, structured scripts, and a shared story queue into one persistent interactive world.
Starts with WASM isolation for untrusted code, then compiles streamed model plans into DAGs that can run in parallel.
Code-first infrastructure connecting GPU and SLURM scheduling, training observability, experiment lineage, and research decisions.
Compiles a TypeScript-like language to vanilla Minecraft datapacks and provides editor, validation, and testing tools around it.
Links page elements to DOM, component, and source locations while keeping final acceptance with the human reviewer.
A spatial-photo experience built around geographic objects, AR scenes, and cloud anchors, connected to a real-time backend.
Compares six latent-space watermarking methods under a shared attack and evaluation protocol, and proposes a low-cost removal attack.
Compiles, executes, limits, and automatically judges multilingual submissions for university ACM/ICPC training, contests, and coursework.
Research, papers, and patent
Efficient execution environments for AI workloads
Current research · AI infrastructure · Runtime / WASM / DAG
Studies how model-generated dynamic tasks can be compiled and executed safely in parallel while controlling isolation boundaries, tool capabilities, scheduling overhead, and failure recovery.
Latent Rules Without Rewards: Unsupervised Environment Identification for Continual RL
Current research · Under review · Unsupervised world-model pretraining
Studies how latent rule representations of environment structure can be learned from interaction trajectories without task labels or rewards, then used for environment identification and transfer.
- Paper
STAM-LSGRU: a spatiotemporal radar echo extrapolation algorithm with edge computing for short-term forecasting
Cheng, H., Cui, M., & Shi, Y. (2024). Journal of Cloud Computing 13, 100. DOI: 10.1186/s13677-024-00660-6.
- Paper
Distributed service caching with deep reinforcement learning for sustainable edge computing in large-scale AI
Liu, W., Bilal, M., Shi, Y., & Xu, X. (2025). Digital Communications and Networks 11(5), 1447–1456. DOI: 10.1016/j.dcan.2024.11.009.
- Patent
A Blockchain Address Identity Inference Method and System Based on Deep Graph Neural Network
Granted Chinese invention patent; co-inventor. The work originated in Shisho Guard’s blockchain address identity inference system.
Competition awards
- First Prize, 17th National College Student Information Security Contest
- First Prize, 15th China College Students Service Outsourcing Innovation and Entrepreneurship Competition
- Second Prize, 16th National College Student Information Security Contest
- First Prize, Jiangsu Selection of the Challenge Cup National College Student Academic and Technology Competition
Skills and languages
- Programming
- TypeScript, Python, Go, C#, C++, CUDA, JavaScript
- AI and systems
- LLM agents, tool use, RAG, PyTorch, WASM/WASI, Linux, Docker, SLURM, AWS Lambda
- Games and tools
- Unity, SceneKit, Minecraft datapacks, compilers and LSPs, automated judging, SSE/WebSocket
- Languages
- English: IELTS 8.0 (Listening 9.0, Reading 9.0, Writing 7.0, Speaking 7.5); Japanese: basic working communication