Engineering
Notes from building and running a small ML platform - a heterogeneous GPU cluster doing real work: inference serving, agentic tooling, Kubernetes operations, and the occasional post-mortem. More about me.
Currently
Now
Breadcrumbs for the human in the loop
A self-hosted agent reviewing every PR on my own GPUs, so agent-written code still gets a second pair of eyes.
Write-up · 3 min→
Before
Losing the hypervisor for Talos
Wiped a 7-node GPU cluster and rebuilt it from k3s-on-Proxmox to bare-metal Talos - a full rebuild on your own metal, no cloud provider.
Build-log · 5 min→
Elsewhere
Code and tools on GitHub
Projects and open-source work live at github.com/imlach.
GitHub ↗→
Writing
2 entries
Breadcrumbs for the human in the loop
Coding agents have made it easier than ever to ship changes, and harder than ever to keep an eye on them all. cora is the review agent I built and run on my own GPUs, leaving breadcrumbs for the human in the loop.
Losing the hypervisor for Talos on bare metal
Wiped a 7-node heterogeneous GPU cluster and rebuilt it from k3s-on-Proxmox to bare-metal Talos. How to do a full rebuild without a cloud provider.