I mess with GPU scheduling and heterogeneous infra on Kubernetes.
How GPUs get carved up, shared, and kept from stepping on each other across different hardware, vendors, and clusters. And what explodes when that goes wrong.
I’ve got a thing for edge compute, bare-metal infra, and accelerator silicon. If there’s a conference with hardware in Asia, I’m probably there - Computex, Semicon Taiwan, SuperAI, or WeSemiBay in Shenzhen.
Right now I’m at Dynamia working on HAMi, a CNCF sandbox project that does GPU virtualization and sharing on Kubernetes. One Helm chart, 11+ GPU vendors, zero vendor lock-in It’s the only CNCF-backed stack that handles heterogeneous compute.
- Making GPU sharing work across NVIDIA, AMD, AWS Trainium/Inferentia, Huawei Ascend, etc.
- Turning GitHub stars into production clusters
- KubeCon talks, partner stuff, helping enterprises and saving neoclouds from vendor-lockin
- Convincing GPU accelerator startups that they should love OSS
I’ve been on the HW/SW boundary for about 30 years. Reverse-engineering codecs, demuxers, Apple’s bootloader back when that was fun. Built semiconductor test rigs. Ran real-time seismic data systems on edge hardware that barely had a pulse. Did healthcare infra, data center construction, CTO roles in startups and giant government orgs.