Corbin Imgrund

LinkedIn GitHub

AI & Automation

I build automation that does real work and stays under my control. Here are two systems I designed, with the calls I made.

Selected work

Local LLM routing on a GEEKOM A9 Max mini PC

  • Role: Owner. I designed the setup, tuned the hardware and run the stack.
  • When: 2026
  • Context: I wanted my agent work to run on my own hardware instead of hosted models. The hard part was fitting one large model and a team of agents onto a mini PC with an integrated GPU.

  • What I did:

    • I set up Ubuntu on the A9 with a BIOS and pin configuration I tuned for local inference. I cut the GPU's fixed memory reservation to 512 MB and raised the shared memory it can borrow to 128 GiB, so the model runs from system memory.
    • I chose Vulkan, with the open-source Mesa RADV driver, over ROCm. I pinned one model for every agent, with no fallback model.
    • After the GPU crashed under agent load, I rolled the GPU settings back and set a rule: no raising them until 48 clean hours and a new benchmark. I picked stability over speed.
  • Artifact: How my local LLM routing works
  • Outcome: As of Oct 2026, every active agent runs on the local model.
  • Stack/tools: GEEKOM A9 Max, AMD Ryzen AI 9 HX 370 with Radeon 890M graphics, 128 GB RAM, Ubuntu, Vulkan (Mesa RADV), Ollama, Qwen 3.8 Flash-Next, Paperclip, Hermes

Built with a team of AI agents under my direction. I set the scope, made the product calls, and nothing shipped without my yes.

Grok Bot: a team of AI agents with one front door

  • Role: Owner. I designed the operation and wrote its rules. Merges, deploys and anything public wait on my yes.
  • When: 2026
  • Context: I needed a lot of agents working at once without losing track of who approved what. Every approval had to live in one place, and nothing risky could happen without me.

  • What I did:

    • I designed the routing so one front door hands work to specialist agents. I talk to one chief-of-staff agent, specialists never ask me for approvals, and every handoff names the next owner and points to a file.
    • I set up one shared approvals ledger. Every yes goes in with my exact words and the time, and an agent acts only on a matching entry that hasn't expired.
    • I made the A9 the first stop. Long, repetitive local work goes to my local stack first, and hosted models take it only when the A9 is offline or the job is urgent.
  • Artifact: How my Grok Bot automation works
  • Outcome: As of Oct 2026, the operation runs 13 agents: a chief-of-staff agent and 12 specialists.
  • Stack/tools: Grok Bot, plus the A9 local stack above

Built with a team of AI agents under my direction. I set the scope, made the product calls, and nothing shipped without my yes.