Joseph Maxwell
- Datacenter Technician at Google
- TPU & GPU Infrastructure
- AI-Assisted Hardware Diagnostics
I turn up and repair Google Cloud TPU systems (TPU v5e to Ironwood) and NVIDIA GPU servers. I test fiber across Optical Circuit Switches, diagnose hardware faults with an AI agent, and maintain an Apps Script tool for my area's logistics requests.
Selected Work
Full experience- 01 TPU and GPU Turn-Up Turn up and maintain Google Cloud TPU systems (TPU v5e to Ironwood) and NVIDIA GPU servers. Once the machines are racked and cabled, QA them and correct mistakes and defects, including on multi-site deployments.
- 02 AI-Assisted Hardware Diagnostics & Mentorship Find the root cause of hardware failures across 23+ architectures, including TPU and NVIDIA GPU systems, reading logs and error signatures with an AI agent's help. Train and mentor other technicians on the same AI workflows.
- 03 Area Logistics Tooling Maintain and extend a logistics request tool for one area of the site, built on a mentor's tracking spreadsheet, with Google Apps Script, Forms and Sheets. A validated intake form means each detail is entered once and fewer go missing between shifts.
Accelerators
The TPU and GPU platforms I work on at Google, and what I do on every one of them.
- TPU v5e
- TPU v5p
- Trillium TPU v6e
- Ironwood TPU7x
- NVIDIA GPU servers
- Rack QA & turn-up
Once the machines are racked and cabled, I QA them, fix any mistakes or defects, and turn the systems up, at my site and on multi-site deployments. When automation hits a snag it can't handle, I work through it with remote deployment engineers.
- Diagnose & repair
I read the logs and error signatures, with an AI agent's help, to find the failed part, then replace it and update the repair docs so the same fault doesn't come back.
- Fiber testing
I clean every connection, measure receive power, and split the circuit with a loopback to find which side of the patch panel or Optical Circuit Switch has the fault.
Arcade
All games- 01 Floor Space Line up a CPU, an SSD and a GPU to build a server, then merge servers up to a datacenter before the floor runs out of room.
- 02 Loopback A fiber link between two machines still has a fault after cleaning. Loop the light back to split it, then swap the bad part. Fix as many links as you can in 90 seconds.
Kit
What I carry on the floor, and what I run at home.
On the floor
- Screwdrivers, bits and torque adapters For swapping CPUs, GPUs, IPUs, NICs and other parts.
- Thermal paste and lint-free wipes To clean off the old paste before the new goes on.
- USB-C to serial cable Console access to a device over its serial port.
- Loopbacks, retroreflectors and cleaners For LC and MPO connectors.
- Visual fault locator (VFL) Red light that shows a break or a tight bend in a jumper.
- Fiber IR card Shows whether infrared light is coming out of a fiber or an optic.
- Laptop Where most of the diagnosis happens: logs, tickets, and an AI agent to read error signatures.
At home
Guests on one Proxmox host
- Xeon E5-1620 v2, 4 cores / 8 threads
- 64 GB DDR3 ECC
- 512 GB SSD
- 16 TB WD Ultrastar HC550
- Home Assistant OS Home automation, in a VM.
- Kepler ◆ An AI ops agent. It reads the lab's state and asks for approval before it changes anything.
- Karaoke Studio ◆ Turns a song into a karaoke video, with AI vocal separation and word-timed lyrics. The heavy work runs on a cloud GPU.
- Uptime Kuma Watches every service.
- Proxmox Backup Server Nightly backups, with an encrypted copy offsite.
- Homepage A dashboard, and the lab's own docs site.
- Pi-hole (secondary) failover Second DNS server and hot-standby DHCP.
- Jellyfin Media server.
- Reverse proxy TLS for internal service names.
- Dev workspace A development desktop in the browser.
Raspberry Pi Zero W
The primary Pi-hole, for DNS and DHCP.
If the Pi stops answering, the secondary takes over DHCP.
Built by me
I declare every guest in OpenTofu and set up the containers with Ansible. Secrets stay encrypted in the repo.
Notes from the lab
I tested the backups with a full cold restore from the offsite copy, and every chunk verified.
Before a firewall change, I arm a five-minute auto-rollback, so a bad rule can't lock me out.
Pi-hole can't split DHCP between two servers, so the standby runs a watchdog and takes over after three missed offers.
The Xeon has no AVX2, so local AI vocal separation crashed or ran slowly. I moved the heavy part to a cloud GPU and kept a local fallback.