What I cannot create, I do not understand.
I'm Zhiyi Cao. I build infrastructure at Tencent, previously at Baidu, DiDi, and Bonree.
For the past decade, most of my time went into network protocols and low-level systems. Those interests grew out of the work itself — solving concrete problems first, then growing curious, then wanting to build something of my own. That impulse became two open-source stacks (quicX and CppNet) and two books on cloud networking and AI coding.
These days my main focus is AI engineering: AI gateways, AI coding systems, and the engineering problems that surface when large models actually ship. If protocols, systems design, or AI engineering interest you, the long-form writing here is a good place to start.
Works
Start here
If you only read three: two on the AI engineering I do now, one on the protocol decade behind it.
LLM traffic breaks the two hidden assumptions of the traditional API gateway — opaque payloads and deterministic behavior — and forces a new infrastructure layer that governs non-deterministic calls.
The Repriced EngineerAs the industry shifts from rule-based to token-based delivery, the capability bundle of software engineers is being repriced: raw coding depreciates while judgment becomes the new scarce asset.
Why We Need QUICAn explanation of the evolution from HTTP/1.x and HTTP/2 to QUIC, analyzing why we need user-space protocols to solve head-of-line blocking, achieve zero-RTT handshake, and support connection migration.