I am Chi-Wei Lee, a Physics and EECS (AI track) double major at National Tsing Hua University, Taiwan. At the HMI Lab, I work on generative models that reconstruct 3D spatial structure from fMRI. I also build open source tools and fix correctness bugs in scientific and GPU software. My research interests include predictive coding, associative memory, and sampling-based inference.
| Project | Why it exists |
|---|---|
| quantum-allosteric-scanner | Predicts protein allosteric sites from one unbound structure with continuous-time quantum walks. |
| agentleaks | Finds and redacts exposed API keys in local coding-agent histories. |
| inference-visually | Interactive, browser-based explanations of LLM inference. Try the live demo. |
| NVIDIA cuDF | GPU dataframe library where I contributed two merged correctness fixes. |
| Stanza | NLP toolkit where I contributed three merged token-reconstruction fixes. |
| Kornia | Computer vision library where I fixed a per-class detection metric. |
| Project | Contribution |
|---|---|
| NVIDIA cuDF | Fixed empty strings becoming null in Series.str.join and time-of-day loss at month boundaries. |
| Stanza | Repaired multi-word token reconstruction across three merged PRs. |
| Kornia | Corrected per-class recall in mean average precision. |
| Burn | Clamped probabilities in label-smoothed cross entropy. |
| pymdp | Fixed label-list indexing to select blocks instead of diagonals. |
| braindecode | Corrected Hilbert-frequency output length and Nyquist handling. |
I focus on bugs that produce wrong answers, with reproducible cases and regression tests.
My research and creative projects include X-Ray, NSF-HDR, MatrixQR, and my bilingual research website. I also build practical local tools: gpuwho, papercompass, installwall, cardsmith, modelshift, llm-doctor, slopblock, snipmd, cliffhanger, songforge, and shiftgear.
I captained a team that placed 2nd of 609 in the 2025 NSF HDR ML Challenge anomaly-detection track. I also visited UCLA Samueli through the Taiwan Global Pathfinders program in 2026. More about my research and experience.
