Yilin Zhang · Gavin Z
Exploring how agents work with people and organizations; finding the constant in change
Agent PM & full-stack builder · INTJ
Who I am
BSc in CS at the University of Toronto, MSc in CS at Tsinghua — never went the algorithm track, ended up in product.
Heavy Claude Code user; I want to build fun things with AI.
What I do
now · Kuse｜Head of Agent
before · Meituan｜Senior Product Manager
What I'm good at
- The well-rounded, AI-fluent kind of PM — user PM, search/reco strategy PM, AI / Agent PM; with Claude Code I found the joy of coding again.
- The algorithms guy who can actually write papers — first-author in IEEE TKDE (CCF-A); the go-to "paper fixer" in school. Fine-tuned GPT-2 back in the day.
- The data analyst with a sense for business — at Meituan the business analysts came to me for SQL; at Kuse I turned company data analysis into an automated loop.
Find me
I mostly post my AI notes and takes on 即刻 (Jike) — follow along. Open to coffee chats — reach me on 即刻 or email.
Work
Kuse · Head of Agent 2025.10–now
An AI Workspace for knowledge workers worldwide. I own Agent end-to-end:
- Built the agent harness from scratch: system prompt, tools, Skills, context management, sandbox
- Vibe-coded a one-line-to-run automated eval + self-iteration loop (open-sourced as agent-evalkit)
- Turned company data analysis into fully automated loops: daily / weekly / monthly
- Rebuilt billing & credits so the product actually makes money
Meituan · Senior Product Manager 2022–2025
- Merchant-side AI digital employee — among the company's first publicly launched AI agents; owned product architecture, took it from demo to production (a business line serving tens of millions of users)
- Consumer AI outbound calls: built model selection, eval sets, and the prompt system from scratch
- S-tier marketplace project: search & growth across 10+ teams
Earlier · Research 2018–2021
- MSc in NLP at Tsinghua; first-author paper in IEEE TKDE (CCF-A)
Writing
The story of KDA is not just about scaling linear attention. It is about choosing a technical direction, betting on it, and staying with it long enough to make it work at 2.8T parameters.
What loop engineering actually is, how to build a good loop, and why a new "X engineering" keeps appearing every few months.
Press
- Business Insider (2026.2) — "How Chinese AI products are built differently": cost efficiency, open source, market constraints. read →
- LatePost AI Agent closed-door session · product representative · Apr 2025
Open source
agent-evalkit GitHub ↗
The open-source cut of the agent-eval framework I built at work — eval engine + a visual dashboard.