AI should interpret Zigbee data, not replace Zigbee rules
So the argument keeps landing on the wrong question: whether a cheap board can run a model, not whether the model belongs near the thing that flips the switch. Start from a Zigbee network, motion, contact, temperature, and power sensors into a coordinator, and most of what happens between a sensor event and an actuator response is a job for deterministic code, not inference. Latency and reliability there favor a state machine over a model call.
the coordinator already knows what to do
Small-board AI collects two flavors of bad advice. One crowd treats a Raspberry Pi or an Orange Pi like a shrunken gaming rig. The other writes the category off because it doesn't feel like a frontier chatbot. Both miss the useful middle: quiet, bounded jobs near the sensors that keep working without a cloud account and barely sip power. The board matters less than the appliance around it: storage, supply, cooling, OS, whatever radio it carries. A script run once from a shell is still several decisions from a household feature.
write the rule before you shop for boards
Write the event, the input, the output, and the deadline in four lines before you shop for hardware. "Add AI to the camera" is not a job. "When a parcel sits in the porch zone for twenty seconds, send one local notification with a thumbnail" is testable, and testing it shows which parts never needed a model. Keep the rule local and instant, let history accumulate somewhere boring, and ask a model only for a bounded recommendation once there's enough history to judge. Test with the actual room, because edge deployments amplify ordinary variation more than any model revision does: evening light, a busy Home Assistant database. Log the board, the model version, expected versus actual action, latency cold and warm, and what a false or missed action looks like, plus its recovery. Accuracy alone won't say whether it survives a reboot or fails safely when inference isn't there, and anything shared with the household needs visible state and a fallback.
a model is not a fuse
The failure I see most often is routing every motion event through an LLM before it's allowed to turn anything on. A model is not a fuse. It's genuinely good at interpreting a fuzzy request or ranking options after the fact. It has no business guarding a lock, a heater, an alarm, a valve, a battery, or a person alone. Ranges, permissions, timeouts, confirmations, and interlocks belong in deterministic code. Treat anything text-shaped that reaches the system, camera captions, calendar entries, a webpage, an MQTT payload, a voice transcript, as data, never as instruction.
local inference doesn't erase the footage
Running the model on your own hardware solves exactly one privacy problem: it stops one upload. It does nothing about the recordings, event logs, thumbnails, backups, or the dashboard someone left open on a phone. You still have to decide what gets collected, how long it sits around, who can pull it up later, and what gets logged while you're debugging at midnight. The most private frame is the one the camera never captures. The next best is discarded right after the decision.
the acceptance test happens at the wall, not the benchmark
Run the thing as an appliance, not a demo you're proud of. Use a service manager or a small, pinned container, keep configuration separate from the model files, and write a health check that proves the function works, not that a process holds a port open. Back up the controller state and keep a known-good image for any node installed somewhere inconvenient. Let it run to thermal equilibrium and measure power at the wall, because a board that throttles, swaps, corrupts storage, or hogs the automation host isn't ready, whatever the demo looked like. Point the model at trends and ambiguity, never the critical path of a button press, and when it's gone the automation should get dumber, not stop.
I've made my peace with what that costs. You don't get the impressive story where the model is the brains of the house, deciding things live and looking clever. You get a house that's mostly dumb switches wired correctly, with a small, slow layer of judgment reading the history nobody has time for. That's the trade I take on purpose: less spectacle at the demo, a porch light that still works the night the model doesn't load.