Developer · Researcher · Mathematician
I build tools that bridge mathematics, AI, and real-world applications — Microsoft Store desktop apps, full Python runtimes for iOS, and on-device IDEs for iPad.
2025 S.T. Yau High School Science Award · Bronze, Grand Finals · Silver, Asia Regional
Published research · peer-reviewed
Limiting spectral distributions of random matrices
Cambridge Journal for Junior Scientists, Volume 3 (2026), Number 2, pp. 365–382. Published 30 July 2026 by International Press of Boston. Open access.
doi:10.4310/CJJS.260626162006 · grew out of the work that took Bronze in Mathematics at the 2025 S.-T. Yau High School Science Award finals
30+ native Python libraries cross-compiled for iOS · 5 packages on PyPI · 1 peer-reviewed paper · 2 science awards
Selected work
A self-contained, offline dev / scientific / AI workstation — Monaco editor, integrated terminal, Python 3.14 + C/C++/Fortran, on-device pdflatex, and local LLMs via llama.cpp.
A full Python runtime for iOS with the first public native PyTorch build, HuggingFace transformers, Rust tokenizers, NumPy/SciPy/scikit-learn, manim, and C/C++/Fortran interpreters. Powers CodeBench.
My Yau-award research shipped as software, twice. EigenDenoise is a Metal-accelerated macOS image denoiser built on random matrix theory; Generalized Covariance Matrix is the Windows research tool for eigenvalue spectral distributions — analyse ESDs, visualise spectral densities and export results, fully offline.
A boot manager for your iPad. It boots real operating systems — 64-bit Linux, classic Windows, graphical desktops — inside a WebKit-hosted emulation stack, with real internet and real multi-core. Everything runs on-device: no servers, no streaming, no jailbreak.
My own tropical-cyclone model: track, central pressure, intensity and the wind-field radii five days out. It reads analysis fields valid at or before the issue time — never a forecast field, and never another agency's track. Live on the tracks explorer over real JMA forecasts and past storms.
A compiler and application builder written with nothing but the Python standard library. It turns Python into real machine code — and where it needs C in the middle, it compiles that C with a C compiler it also wrote. No Cython, Nuitka, Rust, gcc or clang at any point.
A static reverse-engineering of Apple's on-device foundation model — the sparse Mixture-of-Experts backbone in Apple Intelligence — decoded from its container formats down to a component-validated PyTorch forward pass. Original research only: no Apple weights or tokenizer are redistributed.
What I do
01
Random matrix theory, the Stieltjes transform, limiting spectral distributions, and statistical modeling.
02
On-device PyTorch, HuggingFace transformers, scikit-learn, and local LLMs running fully offline.
03
Windows Store and App Store apps, native iOS frameworks, and open-source packages on PyPI.
04
Math animation with Manim, hardware video encoding with FFmpeg, and GPU 2D graphics with Cairo on Metal.