Hi, my name is Run Zhang.

你好，我是张润 — welcome to my page.

Portrait of Run Zhang

About

I'm a software engineer working on AI. Most of what I do sits where machine learning meets real systems — training and evaluating models in Python and PyTorch, then making them run reliably and fast once they leave the notebook.

I've built perception for autonomous vehicles, researched graph neural networks for text, and worked on video prediction. Lately I've been going deeper on large language models and the engineering around them: retrieval, agents, evaluation, and CUDA for squeezing more out of the hardware. I'm drawn to problems that are both useful and genuinely hard.

Education

  • MS Information Systems
    New York University
    September 2023 – May 2025
  • MSc Artificial Intelligence
    University of Southampton
    September 2019 – October 2020
  • BEng Software Engineering
    Guangzhou University
    September 2015 – July 2019

Projects

Realistic 3D Unreal Engine game environment scene

Realistic 3D Unreal Engine Game Environment

Built a realistic, interactive 3D game scene in Unreal Engine with an AI-assisted asset pipeline, generating 3D meshes using Tripo3D and scripting core gameplay mechanics and environmental interactions using Claude.

  • Unreal Engine
  • Tripo3D
  • Generative AI
  • Game Dev
Custom physical bipedal walking robot standing and walking

Custom Bipedal Walking Robot

Designed and built a physical standing and walking bipedal humanoid robot from scratch, spanning mechanical CAD in Onshape, custom PCB design, reinforcement learning locomotion policies trained in Isaac Sim and MuJoCo, and ROS 2 real-time motor control.

  • Robotics
  • Reinforcement Learning
  • Isaac Sim
  • ROS 2
  • Hardware
End-to-end local decision model running automation on macOS

End-to-End Local Decision Model for macOS

Trained a compact, specialized local decision model for macOS desktop automation workflows, building the entire machine learning pipeline independently—from custom tokenizer training and pre-training to fine-tuning and on-device deployment optimized for Apple Silicon.

  • On-Device AI
  • Apple Silicon
  • LLMs
  • Tokenizer Training
  • macOS
Neural network controller steering an autonomous surface vehicle

Autonomous Surface Vehicle Controller

My Master's dissertation focused on implementing Evolutionary Algorithms to develop a vehicle controller, which selected actions based on the current state of the environment or the vehicle itself. Additionally, the project explored deep reinforcement learning and compared the two approaches.

  • Evolutionary Algorithms
  • Deep RL
  • C++
Relational graph generated from text data

Graph Attention Networks with automatic graph generation on text data

While working with text data, I realized that structured data is commonly represented using relational graphs, where two nodes are connected if they are related. This observation naturally led to the idea of exploring whether deep learning models can be trained to generate these relational graphs.

  • Graph Neural Networks
  • LSTM
  • NLP
SimVP architecture for video prediction

Predicting the future semantic masks of moving objects

For the deep learning course at NYU, I worked with two classmates to implement the open-source SimVP model on a video semantic prediction task, then explored VQ-VAE pre-training as a follow-up.

  • Video Prediction
  • Segmentation
  • PyTorch