Accurate and Efficient CUDA Generation through Reinforcement Learning (2025) Research

Tech: Python, PyTorch, Triton, CUDA, Reinforcement Learning (GRPO)

A reinforcement learning framework for generating functionally correct and hardware-efficient GPU kernels. This project employs a compiler-inspired multi-stage architecture that decouples correctness validation from performance optimization, using GRPO for policy optimization. The system rewards compilability, functional correctness, and runtime speed, enabling iterative refinement of Triton kernels for efficient GPU execution. Developed for ECE 1508: Reinforcement Learning course at University of Toronto.

Provenance of Training without Training Data: Privacy-Preserving DNN Model Ownership Verification (2023) Research

Tech: Python, PyTorch, Deep Learning, Model Watermarking

A novel intellectual property protection mechanism for deep neural networks that enables model ownership verification without requiring access to training data. This work addresses the challenge of protecting DNN models against model extraction and unauthorized reuse by designing robust watermarking techniques that preserve privacy while ensuring reliable ownership verification. Published at WWW '23.

HMARK: Radioactive Multi-Bit Semantic-Latent Watermarking for Diffusion Models (2025) Research

Tech: Python, PyTorch, Diffusion Models

HMARK is a novel watermarking technique for diffusion models that embeds multi-bit information into generated images while maintaining image quality. This work addresses intellectual property protection and detection of AI-generated content by creating robust, semantic-latent watermarks that are resistant to removal attacks.

HarmonicAttack: An Adaptive Cross-Domain Audio Watermark Removal (2025) Research

Tech: Python, PyTorch, Audio Processing

HarmonicAttack is an adaptive attack framework designed to evaluate the robustness of audio watermarking systems. This research investigates cross-domain watermark removal techniques and helps improve the security of audio watermarking by identifying vulnerabilities in existing approaches.

LDPKiT: Superimposing Remote Queries for Privacy-Preserving Local Model Training (2025) Research

Tech: Python, PyTorch, Differential Privacy

LDPKiT introduces a novel approach to privacy-preserving machine learning by superimposing remote queries during local model training. This work focuses on improving model utility under local differential privacy (LDP) constraints while maintaining strong privacy guarantees, addressing the fundamental trade-off between privacy and accuracy in federated learning scenarios.

PixelArt (2022) Project

Tech: Python 3, Blender

It is undergrad 4th-year capstone project. We developed a general algorithm for synthetic static graphical data generation that accepts valid 3D objects. We also developed neural networks for object detection and lens blur effect. We then verified the quality of generated dataset on lens blur image processing application. Our project was one of the top projects that year and got the Certificate of Distinction.

RECTNet (2020) Project

Tech: Python 3

It is an emotion detection AI model developed by Rain, Evan, Cassie and Tim (RECT). We developed multiple AI models for face detection and emotion detection with the technique of transfer learning. The Web APP is available at http://rectnet.ml/demo. We used Affectnet as our database to train our models.

LegUp High-Level Synthesis (2019) Research

Tech: C++, Python, LLVM, FPGA

Enhanced the LegUp High-Level Synthesis framework to better exploit register-rich FPGA architectures. Implemented LLVM backend extensions that improved pipeline depth and performance of synthesized FPGA circuits. Upgraded LegUp to include Stratix10 board support and optimized algorithms to reduce wall-clock time by taking advantage of the register-rich routing architecture. Evaluated research prototypes on Intel Stratix 10 FPGAs to validate architectural and compiler-level optimizations. Published at ICFPT 2019.

Supercalifragilisticexpialidocious Mapping Service (2019)Project

Tech: C++

It is an integrated project of map design. It is a team project accomplished by a group of three. The project is entirely programmed in C++ from UI design to back-end programming. The project incorporates the use of gtk, libcurl, EZGL for graphics and data. The use of algorithms such as A*, greedy, k-opt, etc. solves classic TSP problem. The features of this map include all the basic features other State of Arts (e.g. Google Maps)have, such as map-changing, IP Geolocation, multi-functional search bar, help button, pop-up window, etc. Our map also inmplemented 3D streetview and 3D tour. The future development would be augmented reality. Project Link

Game of Life (2019)Project

Tech: C & FPGA

This is a team project accomplished by the team of two. The project collaborates with the use of FPGA keys and PS2 Keyboard. The player can choose the game mode, observe the survival of germs, start again or end the game. This is a simple game using the knowledge of polling and interrupt, as well as some software programming logic. It indicates a starting point of embedded programming.

Music Game Integrated with FPGA (2018) Project

Tech: Verilog & FPGA

The design is a combination of five parts, namely the video input, PS2 keyboard input, game logic, VGA display output, and audio output. Video input and VGA display output interact directly with the game logic, video input provides the real position and VGA’s finite state machine (FSM) logic provides the expected position, game logic compares them and updates the score. Then it outputs a signal to the VGA’s FSM and displays the score on the screen. The PS2 keyboard input interacts with both VGA output and audio output, controlling the speed of the game through Rate Divider and select a song to be played. Project Link

WeChat Mini Program (2018) Project

Tech: Java, HTML, CSS, JavaScript

This is a program that allows users to share videos on the social media. Users can create their own account and password. They can upload their profile pictures and videos. More functions are waiting to be updated.

Small-scaled Python project to fetch information from websites (2018)Project

Tech: Python, Regular Expression

This is a program that can fetch the name of anchors and number of audience from a website's HTML script using web scraper. The principle of this program is the combination of regular expression and Python, which simplifies the part of data analysis. After analyzing the data, it is able to get the popularity of the anchors.

Mann Museum design team with professor Steve Mann (2018)Project

Tech: HTML

It is a research team that does research on professor Mann's work, such as virtual reality, High Dynamic Range(HDR), etc. and proposes ideas on the construction of Mann's Museum in Shenzhen, China.

Method of delivering a learning program to Haiti (2018)Project

As a team, we initiated a pilot program to deliver Khan Academy to Haiti for Haitian NGO Centre d’Inspiration Jeunesse. Since I was the project manager, it was my responsibility to communicate amongst team members and the client. In the end, we proposed a new energy system that works in conditions without electricity or light.

TO BE CONTINUED...