Hi, I am Joe from Taiwan, currently exploring new things in Tokyo.

Most recently I was an Applied Scientist at Amazon, where I worked on improving Kiro coding agents. I used live-session trajectory data to find signals for improving agents through harness optimization and RL training (blog post on the harness side).

I’m broadly interested in retrieval systems and making them faster. On the side I contribute to next-plaid and fast-plaid (asymmetric binary quantization, fused CUDA kernels, faster shortlist pipelines) and built nano-plaid, a from-scratch ColBERT+PLAID implementation in one NumPy file with a Rust SIMD kernel ladder.

I did my PhD in Computer Science at the University of Chicago in the Chicago Human+AI (CHAI) lab, advised by Prof. Chenhao Tan, where I studied how NLP on large corpora can support human decision-making.

News

Date	Event
2026-09	Our ModernBERT-base-msmarco checkpoint became the most downloaded ModernBERT-base embedding model on Hugging Face (293k+ downloads)!
2026-07	Two papers accepted: CodeScout at Findings of ACL 2026 and CodeStruct at ACL 2026!
2026-02	Sharing our Kiro post on learning from live agent trajectories to improve coding agents!
2025-03	Happy to share my attempt at building a multi-GPU embedding server with caching!
2024-12	New post on reproducibility of ModernBERT for retrieval tasks is out!
2024-12	Officially graduated and obtained my PhD degree!