Shoujun (Adrian) Diao

At Lovart I am the backend architect for an AI-agent product: the service topology, the gateways, the model and tool access layers, and the cost of inference across GPUs and vendor APIs. Recent work: moving core services from Python to Go, a 30% TTFT cut that let a 14-node H200 cluster run on 6, and routing with fallback across 10+ model vendors.

Before that, seven years at Tencent: edge computing and AI cameras at YouTu Lab, then tech lead for the gateway, AutoML and data-processing modules of Tencent Cloud's TI platform, which we built from zero to commercialization. Its AI gateway took a 100× traffic surge during the DeepSeek boom without an incident.

I think about inference the way I once thought about Go: position first, then tempo. Most of what I write here is about that.