I build the inference infrastructure behind generative-AI products, and I like explaining systems well enough that other people can see them.
Now · Backend architect at Lovart, an AI-agent product: service topology, gateways, the model and tool access layers, and what inference costs. Recent work: core services Python → Go, a 30% TTFT cut that let a 14-node H200 cluster run on 6, and routing with fallback across 10+ model vendors.
Before · Nine years at Tencent (2016–2025). Five of them as tech lead for the gateway, AutoML and data-processing modules of Tencent Cloud's TI platform, from zero to commercialization; its AI gateway took a 100× traffic surge during the DeepSeek boom without an incident. Before that, edge computing and AI cameras at YouTu Lab, and the brand-advertising platform.
Writing · diaosj.info. In progress: A model gateway is a routing problem, not a proxy · Burn-down: a cost model for inference · Leaving one cloud: notes from a multi-cloud migration.
| Infra | Explaining systems | Site |
|---|---|---|
| p2c-simulator · 5k QPS across 300 rate-limited nodes; why power-of-two-choices beats a global scan | https-simulator · TLS handshake, certificates and MITM, animated | my-garden · source of diaosj.info, Astro |
| concurrency-arena · process vs thread vs event loop vs goroutine under I/O and CPU load |
Elsewhere: speaker in the Apache APISIX community; co-translator of Problem Solving with Algorithms and Data Structures Using Python (2019) and translator of Practical Data Analysis (2018); self-taught Go (weiqi) player, amateur 5-dan.
shoujun.diao@gmail.com · Shanghai