About
I'm Jiejing Zhang. I started AI systems work at Alibaba iDST / DAMO Academy in 2017, putting models onto cloud services, GPUs and NPUs at the edge, real-time video systems, conference products, and tiny embedded hardware.
From 2020 to January 2025, I led an inference-engine team for large-scale foundation-model workloads.
That range taught me one habit: do not trust the datasheet. Measure the machine, find the actual roof, move the work, and measure again.
Tempo9 applies that habit to Apple Silicon. A modern Mac has a fast GPU, a Neural Engine, and unified memory; local inference should use all of them, and it should schedule agent workloads like a server rather than a single prompt box.
I'm based in San Jose, California.
ThinkSpread is independent work, built on my own time and not affiliated with, sponsored by, or endorsed by my employer. Its first project is Tempo9: server-style inference for local AI on a Mac you own.
That is also why every performance claim here ships with its script, its raw data, and its error bars, including the runs that disproved something I believed. A number you cannot reproduce is a claim, not a result.
Contact — jiejing@thinkspread.com · GitHub
Colophon
The mark is a metronome — the NIKKO Wooden Mini Advance on my desk, which quietly named the project before I noticed. My guitar cannot quite reach its 208 bpm prestissimo roofline; my systems work is better at chasing rooflines.