I taught band for 32 years. Every concert we played was rehearsed for weeks, wrong notes and all, before anyone heard it. That is how my wife and I run our house in rural Georgia, too.
Our home runs on Home Assistant, kept local wherever we can: about 500 devices, cameras, sensors, and a few hundred automations. I build it with an AI assistant (Claude) as a working partner. Nothing goes live in our house until it has been tested, watched for a week, and argued about.
Rehearsal Notes from Bell House is a series about what we built, why, and what we got wrong on the way. Post #1 is the mailbox, because it is finished, it works, and everyone has one.
The problem: a mailbox that cried wolf
Our mailbox sits at the end of a long driveway, right on the highway. We had a vibration sensor in it that pushed “Mail may have been delivered” to both our phones whenever the box moved.
The trouble is that a mailbox moves for a lot of reasons. Wind moves it. A sprinkler hits it. We walk down and get the mail ourselves. In one week of watching, 3 of 8 “mail delivered” alerts were not deliveries at all, and all three came on the same Sunday.
Two more things bothered me. The sensor sits at the edge of its radio range, so it drops out now and then. And it reports through its maker’s hub and cloud, so if the internet goes down, the mailbox goes quiet.
What we built: let the camera look
The idea came from watching the footage myself. I could see the carrier’s car stopped at the mailbox without zooming in. So the question was never “did the box move?” It was “is a vehicle stopped at the end of the driveway?”
Here is what happens now when the mailbox moves:
- Two UniFi cameras that already look down the driveway each take three pictures, 3 seconds apart. Six pictures, not one, so a car driving past does not look like a car stopped.
- Home Assistant sends the six pictures to Claude (Anthropic’s AI) through Home Assistant’s AI Task feature, with plain instructions: is a vehicle stopped at the mailbox, how sure are you, and what do you see, including any person on foot and what they are wearing.
- The answer comes back as a verdict: “mail stop” or “no vehicle,” a confidence level, and one sentence of description.
- Both our phones get a push with that sentence and one of the pictures attached. A long-press on the notification shows the full picture. My wife’s verdict: she loves seeing the carrier’s car right there in the notification.
- Nothing pushes between 9 PM and 7 AM.
The whole thing takes 40 seconds to a little over 2 minutes from the box moving to the phone buzzing. For a mailbox, that is plenty fast.
The wrong notes, and how we fixed them
This is the rehearsal part. We ran the camera check for a full week next to the old alert before trusting it.
- It saw a car that wasn’t there. At sunset on the first Sunday, the AI said a “dark-colored vehicle” was stopped at the mailbox in all six photos, with high confidence. Both phones buzzed. I walked down: no mail. The footage showed no vehicle, just long pine shadows across the end of the driveway. We switched from the small, fast AI model to the top one with its thinking turned all the way up. It is slower, about 1 to 2 minutes instead of 20 seconds, and in the week that followed it never invented a vehicle again.
- It got the details wrong. Early on it called a silver SUV “dark” and left out the carrier standing beside it. We rewrote the instructions: always mention a person on foot, describe the vehicle’s real color, and only say “high confidence” when the same vehicle sits in the same spot in at least two pictures.
- It couldn’t tell the mail carrier from us. When one of us parked at the end of the driveway with the car door open, the AI correctly said “a vehicle is stopped,” which read on the phone as a mail delivery. The AI answered the question we asked. We had asked the wrong question for that case, so we closed that gap before the camera check replaced the old alert.
- The sprinklers kept checking the mail. One sprinkler zone hits the mailbox post at night, so the box “moved” on watering nights. The check now ignores movement while that zone is running.
- The assistant’s assumption was wrong, too. While judging that Sunday false alarm, it assumed “no mail on Sundays” and used that as evidence. Our mail runs on Sundays. We threw that rule out: judge the pictures, not the calendar.
Where it stands, and the next rehearsal
Since October 1 the camera check has been our only mail alert. The old “mail may have been delivered” push is retired. When the phone says the mail came, it shows us the car that brought it.
One piece still leans on the cloud: the sensor in the box that starts it all. So today I put an Aqara P100 sensor on the ceiling inside the mailbox, joined straight to our own Zigbee network, no hub and no cloud. It reached the house from the road on the first try, with a solid signal. For the next week it runs side by side with the old sensor, writing down every bump it feels. If it catches every delivery without crying wolf at the wind, the old sensor and its hub come out. That will be a future note.
Your turn
How have you solved mailbox detection? A sensor, a camera, a smart mailbox, or the mail carrier’s own notices? And if you use AI on your camera pictures, how do you keep it from seeing cars in the shadows? I would love to hear what worked for you, and what didn’t.
Next in the series: whatever we finish rehearsing next at Bell House.