Projects

Compiler-Induced Numerical Drift & Backdoor Detection

Research Assistant on Trustworthy AI · Columbia University · Jul 2026 – Sep 2026

Studying whether model compilers (torch.compile, ONNX Runtime) introduce tiny numerical differences that can reveal backdoor attacks on neural networks. Trained a 206K-parameter CNN (85% test accuracy on CIFAR-10) and built a PyTorch harness that compares compiled and uncompiled model outputs bit by bit. Traced the drift to ONNX Runtime's compute kernels rather than its graph optimizations: it starts in the first conv layer and grows 37× by the last layer, without changing the final prediction.

Nova Books — Full-Stack E-Commerce Application

Built a full-stack e-commerce application with product browsing, account management, and purchase tracking, using a Spring Boot backend, React frontend, and MySQL database. Designed and consumed REST APIs for user, product, and purchase CRUD operations, and implemented JWT authentication and role-based authorization with Spring Security for login, registration, and admin-only product management. On the frontend, used React Router route guards, Redux Toolkit for shared auth state, and Axios for API calls, then styled the whole thing as a pre-loved bookstore.

View repository ↗

Cryptocurrency & Options Portfolio Risk Research

Worked as a research assistant under the supervision of Professor Yeguang Chi. Built and cleaned an empirical dataset from multiple exchanges, reconciled inconsistent price quotes, and filtered suspicious observations. Analyzed options data in Python with pandas and evaluated how noisy or inconsistent inputs affected backtest performance.