I commission, validate, and automate bare-metal GPU clusters — from PXE boot and BMC/Redfish, through GPU burn-in and NCCL/InfiniBand validation, up to inference serving. The work most teams never get to: catching lemon nodes before they kill a training run, and debugging the firmware / PCIe / fabric-level failures that don't show up in any dashboard.
A100 · H100 · B200 · RTX 5090
- 🌍 South Florida · remote
- 🔧 GPU bare-metal lifecycle — provisioning → validation → networking → serving
- 📬 fabriciopolicarpo0@gmail.com
GPU cluster commissioning & validation — burn-in and acceptance suites (DCGM diag, gpu-burn, nccl-tests, sample training jobs), node health gating, lemon-node detection
Bare-metal automation — PXE / iPXE, Redfish / IPMI, Kea DHCP, NetBox, Ansible, cloud-init, Nomad — full lifecycle from rack to ready
Fabric & networking — InfiniBand / RoCE, ConnectX-7, rail-optimized topology, NCCL tuning, the kind of OEM misconfigs that get blamed on "the network"
Serving — vLLM, fp8/fp16 config, multi-tenant GPU provisioning on idle compute
Also: PXE/iPXE · Redfish/IPMI · Kea DHCP · NetBox · cloud-init · PostgreSQL · MongoDB · Node.js · NestJS · GraphQL · React · TypeScript
Chidori — GPU-over-IP platform. Remote GPU access over network fabric — low-level CUDA and systems work, TCP-tuned data path. A study in what it takes to move GPU workloads across the wire.
📝 Writing on GPU node burn-in, acceptance testing, and why LINPACK lies to you — in progress.
