Dots Connected
I build software that learns from how people actually use it. The interesting part was never the model or the infra on its own — it's the loop between what people need and a system reliable enough to benefit them. Right now I'm researching multi-agent for personalized learning at the Beckman Institute, and helping build an AI copilot for aircraft. Before this, I built data infrastructure at Sonic SVM on Solana, and worked on growth and ads at Tencent and Baidu. Looking back, the dots connect: each role taught me a different part of that loop — what people need, how to capture the signal, how to make the system hold in production.
I believe a great tool is a bicycle for the mind that amplifies human intelligence.
Thoughts
The Shutter Is Proof You Were There
On photography, presence, and the one thing a perfect AI image can't give back.
Your Agent's Bug Is Structural, Not Runtime
Most teams add another judge when they should change the structure. Two layers of agent reliability, and why we invest in the wrong one.
Most Production Features Don't Need an Agent
Why workflows still beat agents for most production LLM features — and the cost test that tells you which to reach for.
Nowhere to Go
On the sanctuary effect, going offline, and what forced stillness gives back in an attention economy.
The Simplest Truths
Why Buffett's simplest principles are the hardest to live by.
Behind the Build: My Tools & Workflow
A craftsman must first sharpen his tools. How I refined my productivity system through dozens of iterations.
The 59-Try Rule
Why high-upside success is worth 59 shots — a mathematical case for low-cost, high-return experiments.
The Zen of A/B Testing
Small-step trial and error and rapid iteration — and why A/B testing is a compass, not a silver bullet.