RXed AI News

The cold-chain alarm: know your freezer failed at 2am, not at 8am

Sensors that wake you when the temperature drifts, and write your food-safety log while they are at it

A freezer that fails at 2am costs you the stock, the next trading day, and the argument with your insurer. The alarm that would have woken you costs about EUR 150, once, and writes your temperature log while it is at it.

The problem

Cold storage fails quietly. A compressor dies on Saturday night, a door is left ajar after the last delivery, a defrost cycle sticks — and nobody finds out until someone opens up on Monday and smells it. Industry write-ups put a walk-in freezer loss at USD 10,000-50,000 of stock depending on what was in it, and report that around 60% of restaurants hit at least one spoilage incident from equipment failure in a year. The built-in alarm on the unit itself is no help: it beeps into an empty kitchen.

The commercial fix is a monitoring subscription. Those work, and they are priced for chains: roughly USD 50-500 per site per month on 2026 US market pricing, plus USD 150-500 per monitored point in hardware, on a contract. For a bakery with two fridges and one freezer, that is a recurring bill defending stock that turns over every week.

There is a second cost hiding behind the first. If you handle food, EU Regulation 852/2004 expects HACCP-based procedures with records "commensurate with the nature and size" of the business, and temperature is the control point everyone gets asked about. Most small kitchens answer that with a clipboard on the wall and two manual readings a day — filled in honestly on good weeks, filled in from memory on bad ones, and worth nothing for the 14 hours a night nobody is there. Digital records are accepted; nothing in the regulation asks for paper.

The architecture

The whole design sits on two decisions. First, the alarm fires on duration, not on a reading. A door held open during a delivery pushes a fridge over its limit for four minutes and that is normal; the same fridge sitting over its limit for half an hour is a failure. Alarm on the instant reading and staff will mute the system inside a week, which is the real way these projects die.

Second, silence is treated as a fault. A flat battery, a probe knocked loose, a Zigbee link dropped behind a metal door — all of them look identical to "everything is fine" if you only watch for high temperatures. So the watchdog also watches for sensors that stop talking, and shouts about those too. This is the single most important line of configuration in the build.

Tool choices — and why

Home Assistant (Apache-2.0, 87.6k stars, actively developed, August 2026) is the watchdog. It is normally sold as home automation, but what it actually is here is a local event engine with a database and a notification system: it stores every reading, holds the duration rules, and sends the alerts. It runs on a small computer in your own building, so an internet outage does not blind it — a monitoring service you cannot reach is a monitoring service you do not have. Its long-term statistics keep the readings for a year and export to CSV from the history screen, which is the compliance half of this build.

Zigbee2MQTT (GPL-3.0, 15.3k stars, release 2.12.1 in July 2026) talks to the sensors. Buying Zigbee sensors and pairing them through their manufacturer's cloud app would work too, right up to the day the vendor changes their terms or retires the product line. Zigbee2MQTT speaks the radio protocol directly through a EUR 25-35 USB stick, supports several thousand device models, and keeps the sensors as yours rather than as accounts.

The sensors must be the probe kind. This is the one hardware detail that decides whether the project works. A sealed sensor placed inside a freezer is a radio inside a metal box, and it drops off the network after a week or two — a well-documented failure in the Home Assistant community. What works is a unit that mounts on the outside of the door with a stainless probe on a cable running in through the seal: the antenna and the battery stay in the warm, only the probe goes cold. The SONOFF SNZB-02LD is one such device, rated -40 to +125 degrees C with a 1.5 m probe; several Tuya-branded equivalents use the same design. Budget EUR 20-40 per cold room.

ntfy (Apache-2.0, 30.5k stars, active May 2026) is the alarm bell. Email and normal push notifications are the wrong channel for something that has to wake you: they arrive silently at 03:00. ntfy delivers high-priority push that can override a muted phone, it is free on the public server and self-hostable if you would rather nothing left the building, and Home Assistant talks to it out of the box. Telegram works as a second channel; do not make it the only one.

ESPHome (GPL-3.0, 11.3k stars, active June 2026) is the alternative if you want mains-powered, wired probes instead of batteries — a small ESP32 board with DS18B20 probes, no battery to die. More work up front, one whole failure mode removed. Worth it for a cold room you cannot easily reach.

What it costs. Software: EUR 0, all four. Hardware, one-off and approximate because prices move: EUR 20-40 per probe sensor, EUR 25-35 for the Zigbee USB stick, and EUR 80-120 for a small computer if you do not already have one running — call it EUR 120-260 for a kitchen with three cold rooms. Running cost after that is the electricity of a device drawing a few watts, so EUR 0-3 a month. There is no per-sensor fee and no contract, which is the entire argument.

Setup outline

1. Install Home Assistant on a small computer that stays on and stays in the building — a Raspberry Pi, a mini PC, or an existing NAS. If you already run any other blueprint in this library on a home server, put it there.
2. Plug in the Zigbee USB stick, add Zigbee2MQTT, and pair one probe sensor at the bench before it goes anywhere near a freezer. You want to see a temperature appear on screen while you can still reach everything.
3. Mount each sensor with the body outside the door and the probe inside, run through the door seal where the gasket already flexes. Fix the probe tip where the product actually sits, not next to the evaporator fan, where it will read colder than the food does.
4. Leave it a full day untouched and look at the graph. Every fridge has a saw-tooth from its own defrost cycles; you need to see yours before you can set a threshold that is not tripped by normal operation.
5. Set the alarm rules per unit: a temperature limit, and a duration — 30 minutes over the line is a sane starting point for a freezer, shorter for a display case that matters more. Add the escalation: still bad 15 minutes later, alert a second person.
6. Add the sensor-silence rule. If a sensor has sent nothing for an hour, alarm. Do this before you trust the system with anything.
7. Point the alerts at ntfy, install the app on the phones that matter, and set that channel to break through Do Not Disturb. Then test it properly: put a probe in a glass of warm water, walk away, and confirm the phone actually screams. Repeat once a month — an alarm you have never heard is a theory.
8. For the record-keeping half, add the daily summary: a message each morning with yesterday's minimum, maximum and any excursion, and a monthly CSV export from the history screen filed with your food-safety paperwork. Show it to your inspector before you throw the clipboard away, and keep the paper log running in parallel for the first month.

Pitfalls — the real ones

A dead sensor looks exactly like a healthy freezer. No reading means no alarm, and a system that has been quiet for three weeks feels reassuring right up to the morning you find out why. Batteries also weaken faster in the cold. The staleness rule from step 6 is not optional, and the battery level of every sensor belongs on the same daily summary as the temperatures.
Alarm fatigue kills the system, not the technology. A threshold set too tight, or with no duration, produces a nightly false alarm during defrost. Two weeks of that and someone silences the app, and now you have hardware in the freezer and no protection. Tune against a real day of data, then leave it alone.
A sealed sensor inside the box will disconnect. Metal walls block the radio; the community threads are full of people who found this out after a fortnight. Probe on a cable, body on the outside. Do not fight this one.
Probe placement decides what you are measuring. Near the fan, it reads the coldest air in the unit and shows you a healthy number while the shelf by the door is thawing. Put it where the product is, and if a unit stores something you genuinely cannot afford to lose, use two.
The alerting path itself can fail silently. Wi-Fi drops, the phone reinstalls, the push token is revoked. That is why the monthly warm-water test exists — you are testing the whole chain, not the sensor.
This is a warning system, not a food-safety programme. It tells you a temperature moved; it does not tell you whether the stock is still safe or what your HACCP plan says to do about it, and it is no substitute for your own procedures. Confirm with your food-safety inspector that the log format works for them before it becomes your only record — digital records are accepted under the regulation, but what they want to see is a local conversation.
Power cuts take out the watchdog too. If the building loses power the fridge is warming and your monitor is off, which is the worst possible moment for it. A EUR 40-70 small UPS for the computer and the network switch buys you the alert about the outage, which is often the more useful alarm.

Verified repos

Home Assistant Core — Apache-2.0, 87.6k stars, active (2026)
Zigbee2MQTT — GPL-3.0, 15.3k stars, release 2.12.1 (July 2026)
ntfy — Apache-2.0, 30.5k stars, active (2026)
ESPHome — GPL-3.0, 11.3k stars, active (2026)