Mechanical & electro-mechanical engineering graduate from NSYSU, Taiwan.
I like making things fast: GPU inference, ML systems, desktop apps and robots.
Open to roles in GPU and ML systems.
Also CUDA, CUTLASS, TensorRT and ONNX
- Making KataGo's CUDA backend faster. My upstream PR speeds up FP16 transformer search by about 13% with no loss in accuracy (#1282).
- Building cross-platform desktop apps with Tauri v2, Rust and React.
KataGo CUDA speedups · C++, CUDA, CUTLASS
Autotuned CUTLASS GEMMs, FP16-accumulate flash attention, vectorized kernels and CUDA graphs. Search is about 13% faster and NN throughput is 16% higher on an RTX 4060 Ti, with the same accuracy as the official build.
katago-transformer-quantization · Python, TensorRT, ONNX
An end-to-end FP8/INT8 post-training quantization study of KataGo's 70M-parameter transformer. INT8 with SmoothQuant raised network throughput by 37%. The write-up explains why FP8 doesn't pay off on GeForce Ada and why the speedup wasn't worth the accuracy cost.
Katraineo · TypeScript, Rust, Tauri
A port of KaTrain, the Go training app with KataGo feedback, from Python/Kivy to Tauri v2 + React. It keeps KaTrain's SGF format and settings, and ships nightly builds for Windows, macOS and Linux.
Voice-controlled robot · Python, C++, Arduino
Spoken commands go through Canary-1B speech recognition on a desktop GPU, then a FastAPI controller on a Jetson, then non-blocking Arduino motor firmware. CI tests the whole stack without hardware.
Taiwan Mahjong solver · Rust, Tauri, React
A desktop app that recommends the best discard in 16-tile Taiwanese mahjong. It computes shanten and effective tiles (ukeire) with a memoized search in Rust.
Math exam grader · TypeScript, Tauri
Grades scanned math exams. It runs OCR locally through Ollama, scores answers with any OpenAI-compatible model, and exports to CSV or XLSX.
- lightvector/KataGo: about 13% faster CUDA transformer inference (#1282, in review)
- elie222/rakazo: run the deployment model through the operator's signed-in Claude Code CLI (#4066)
- langgenius/dify: Traditional Chinese localization (#3899)
- RooCodeInc/Roo-Code: separate Traditional and Simplified Chinese (#197)
- sstklen/trump-code: PyTorch GPU acceleration and a bitmask-vectorized search (#6, #7)
- Smaller fixes in mlc-ai/mlc-llm, mlc-ai/llm-perf-bench, rust-tw/book-tw, getuka/japanese-parler-tts and aszx87410/japan-taiwan-tax-compare