An Elmer, in ham speak, is a mentor — the experienced operator who answers your dumb questions and gets you on the air. I’m lucky to have real ones. But I’m also an AI engineer with a rack of hardware running large language models in my house, so I’m running an experiment: can a local LLM be a useful second Elmer — a patient, always-available tutor for the Extra exam and the bench?

This post is a work-in-progress log, not a finished how-to. Here’s the real setup and where I’m taking it.

The hardware is already here

I don’t rent AI from anyone. The models run on gear I own and control:

  • A Jetson Orin Nano running a llama.cpp OpenAI-compatible server — Qwen3-8B quantized, a few gigabytes, answering on my LAN.
  • A bigger home AI box (an unlocked AMD BC-250 I’ve written about on my personal blog) running larger Ollama models for the heavier lifting.

Both are offline-capable. No API keys, no per-token bill, no data leaving the house. For a ham, that last part isn’t a privacy nicety — it’s the whole point.

Why local AI fits ham radio specifically

Amateur radio’s entire culture is built around working when the infrastructure doesn’t. An Elmer that lives in someone’s cloud and dies with your internet connection is philosophically backwards for this hobby. A study assistant — or, down the road, a field assistant — that runs on a Pi-class box off a battery is exactly the kind of self-reliant tool ham radio is supposed to celebrate.

Same reason I run the Ham Rack off a drill battery: if it needs the grid or the cloud to work, it’s not really field gear.

What I’m building toward

The experiment, in phases:

  1. Extra tutor (now). Feed the question pool and good reference material into a local model and have it explain concepts — Smith charts, filter math — rather than just quiz me. Retrieval over a curated corpus so it cites real theory, not hallucinated physics.
  2. Bench companion (next). A model that can talk me through a measurement or a build step at the workbench, hands-free.
  3. Field assistant (someday). Something small and battery-friendly that rides along on a POTA activation.

The honest caveats

Two things I’m watching closely, and you should too if you try this:

  • LLMs confidently invent physics. For a study tool that is a serious problem. The fix is grounding the model in real references and treating anything un-cited as suspect — never letting it freelance an equation.
  • Small local models have limits. An 8B model on a Jetson is remarkable, but it’s not a frontier model. Part of this experiment is finding where the local, private, offline tradeoff is worth it — and where it isn’t.

I’ll report real results as the tutor comes together: what it gets right, where it lies, and whether a homebrew AI Elmer actually helps me pass Extra. Follow the Extra series for the study side.

73, W3MRB