Raspberry Pi Lab

Most of my projects sit on top of a single piece of infrastructure: a Raspberry Pi running 24/7 as the always-on layer that handles the slow, free, supervised work my Mac can’t always be awake for. It’s not a server in the production sense — it’s a personal automation lab, and it’s becoming the spine of how the other projects compound.

The discipline is hybrid routing. Slow, free, deterministic work (lint passes, summaries, transcription) goes to the Pi. Fast, high-stakes work (interactive coding, the final review) stays with Claude Code on the Mac. The local model is an advisor; Claude Code verifies it and has the final say — a consensus model, not blind trust.

The infrastructure currently runs three loops, with more queued:

• hearth-overnight. A scheduled lint job that runs at 02:00 nightly. The Mac pushes a documentation snapshot to a private GitHub repo before bed; the Pi pulls it, runs a deterministic consistency pass across the docs, commits a report, and opens a pull request. Mornings start with /pi — Claude verifies the findings, posts a consensus review, and merges. No supervision required; the work is done before I wake up.

• Family voicemail / IVR. A Twilio-backed voicemail system on a dedicated phone number. Family and close friends call in, press 1 to leave a message, and I’m notified — recordings stored locally on the Pi, with optional on-device transcription via Whisper. Private, self-hosted, no cloud transcription. (In planning; requirements doc drafted.)

• Local LLM second brain. A 7-billion-parameter coder model (qwen2.5-coder:7b) running on Ollama against the home network. Used for code cross-checks, longer reasoning passes that don’t need to round-trip to the cloud, and as a sanity check when Claude Code wants a second opinion. (Currently hosted on the M2 Mac — the Pi 4 has no NPU and CPU-only inference of a 7B model is too slow for interactive use; revisiting if always-on phone-reachable access becomes the goal.)

The deeper idea sitting under all of it is loop engineering — the practice of designing recursive Plan → Execute → Observe → Adapt cycles that prompt agents on their own, rather than writing one-shot prompts manually. The Pi is where those loops live when they don’t need a human in the loop in real time.