Watchman Your AI watchman.
Watchman makes sensors smart. A thermometer that beeps at eight degrees becomes one that can tell a door left open from a compressor losing the fight, and warns you before the room is warm. Temperature, sound and sight, on the sensors you already have or on ours. It learns what normal looks and sounds like for each thing it watches, tells you when that changes, and answers questions on WhatsApp in Hebrew or English.
No connection to your PLC, your PC or your control network. No software to install. No video. For the most demanding sites, everything runs on premises.
Running today on ESP32 temperature nodes, a Raspberry Pi with a four-microphone array, and a people-counting camera that runs its own detector.
- Temp · cold room4.2 °C, normal for this hour
- Mic · compressorrunning 71 % of the last hour, +40 % vs last Tuesday
- Door · cold roomno data since 09:12, device offlineabstains
- Camera · doorway3 crossings since 13:00, quiet
Compressor working harder. Room still cold. Door unknown.
Check the seal and the condenser. Door sensor offline since 09:12.
Your sensors measure. Nobody understands them.
A sensor that beeps at a threshold is a smoke detector for a problem that has already happened.
Most small plants already have sensors: a logger in the cold room, a thermostat on the oven, a buzzer on the compressor, maybe an app on someone's phone. Each one knows a number. None of them knows what the number means, what it was yesterday at this hour, or what the sensor next to it is saying. So the alarm comes late, comes wrong, or comes to nobody.
The alarm that comes too late
A dirty condenser, a tired door seal or low refrigerant make the compressor run harder for days before the room warms. Temperature is the last number to move, and it is the one with the alarm on it.
The alarm that cries wolf
A door opening, a defrost cycle, a half-degree sensor wobbling across a limit. Each one fires the same beep. After a month, nobody reacts to it.
The app nobody opens
Your floor works from a phone, in a glove, in Hebrew. A dashboard behind a login is something nobody on shift will ever look at, and at three in the morning nobody is looking at anything.
It does the rounds. Continuously.
Connect what you have, add what you need
Sensors you already own join, whatever they speak: if they don't talk MQTT, we build the gateway or the network they need. Ours stick on: a temperature node, a microphone near the compressor, a camera on the doorway, a plug that measures power. USB power or a battery, and Wi-Fi to the internet. Ours need no electrician and no cabinet work, and nothing goes on your control network. A small site can be running in an afternoon. A large one, we plan with you.
It learns what it is watching
Tell Watchman what each sensor is attached to, and it knows what to work out. A thermometer in a freezer is not a thermometer in an oven. Then it learns normal for that thing: this sensor, at this hour of the day, and Planned in this weather. A compressor louder than it was last Tuesday afternoon is worth a message. A cold room warmer at 14:00 than at 04:00 is not.
Ask it, or let it call you
Questions go to a WhatsApp number, typed or spoken, in Hebrew or English: “what is the temperature in the cold room”, “show me a graph of the compressor since Sunday”, “is anything wrong”. Alerts arrive on the phones or the group you name, with the reading, how old it is, and a message when it recovers.
Every answer carries a unit, a timestamp and an accuracy. A sensor accurate to half a degree is quoted to half a degree, never to the fourth decimal it happens to output.
One thermometer. Five answers.
The sensor is the smallest part of Watchman. The rest is what Watchman knows about the thing the sensor is attached to.
| You install | Watchman also tells you |
|---|---|
| A thermometer in a cold room | Door openings, defrost cycles and how long they take, the temperature the food actually feels, how fast the room leaks heat, how strong the cooling still is Planned |
| A microphone near a compressor | Running or stopped, duty cycle, short-cycling, speed, fault buzzers beeping at an empty room In pilot |
| A door contact | Door left open, openings per hour, and which warm-ups it explains Planned |
| A plug that measures power | Energy, cost per month, run state, and a check on everything the microphone claims Planned |
| A camera on a doorway | People counted and line crossings, and one still on request. Reading a gauge or a display Planned |
Alarms follow what the food feels, not what the air does.
Air in a cold room jumps every time the door opens. The food barely notices. Watchman alarms on the food's temperature, worked out from the air, so a door opening is not an alarm and a room that is really warming is. A defrost cycle is expected, and a defrost that takes twice as long as usual is worth a message. Planned
You don't configure it. You tell it.
Tell me if the cold room goes over 8 degrees for five minutes.
That sentence creates the alert, and Watchman reads it back so you can check it.
This thermometer is in the walk-in freezer, and the microphone by the compressor belongs to it.
That sentence tells Watchman what it is watching, and draws the map it reasons on.
Report every ten seconds on the bench sensor.
That sentence changes the device's own reporting interval, and the device confirms it did.
Change your mind at 03:00 from your phone, in your own words. There is a dashboard for the supervisor who wants one. Nobody on the floor needs it.
A microphone in a workplace is a serious thing. Here is exactly what leaves the room.
Every vendor says “we don't record”. We think you should be able to check.
Features, not audio.
The microphone runs continuously, because an impact lasts under a tenth of a second and a device that wakes up once a minute would miss it. What it sends is numbers: a level, eight frequency bands, whether the machine is running, its speed. A few numbers a second, a thousand times smaller than the sound, and they cannot be turned back into speech.
Rows, not recordings.
The numbers land in a database next to the temperatures. A raw clip is stored only on an event, only for a few seconds, only with an expiry, and only if that device's clip flag is on. It ships off.
It reads a picture, not a sound.
When you ask Watchman to listen, it asks the device for a short clip, renders a spectrogram in the cloud, and the model reads the image. The audio stays yours, for a human to play.
The same rule holds for the camera: detections, and one still on event or on request. Never a video stream, never continuous recording.
Or keep all of it on premises.
For the most demanding sites, the whole system runs inside your plant: the broker, the database, the rules and the model, on your network. The three checks below work exactly the same.
What it is not. Watchman is a notification, not a protection system. It does not stop machines, it does not replace a safety interlock, a fire panel or an alarm system, and we will not describe it as one.
Three checks your own engineer can run
- V1 · Read the rowsEvery value the model has ever seen is a row in the database. Open it.
- V2 · Watch the wireMirror a device's traffic. Message sizes show numbers, not audio.
- V3 · Flip the flagThe per-device clip flag is a column. With it off, the clip topic never carries a byte.
Every sensor keeps its own line. A quiet one abstains.
Watchman never averages its sensors into one number. Each keeps its own row and its own reading. Agreement is shown, not assumed. A sensor that has gone quiet abstains in grey: it never counts as “normal” and it never counts as “breaching”. The verdict follows from the rows above it, and every row is in the database.
“No data” is an answer, not a zero. A dead sensor reported as “fine” is how plants learn to ignore alerts.
- Temp · cold room4.2 °C, normal for this hour
- Mic · compressorrunning 71 % of the last hour, +40 % vs last Tuesday
- Door · cold roomno data since 09:12, device offlineabstains
- Camera · doorway3 crossings since 13:00, quiet
Compressor working harder. Room still cold. Door unknown.
Check the seal and the condenser. Door sensor offline since 09:12.
Under the hood. We insist you look.
You are going to audit this anyway, so here is what Watchman actually does, in the language of the people who will be doing the auditing.
The model explains what the sensors saw. It never decides what to do.
The broker, the database and a rules loop handle real time. Alerts are evaluated every ten seconds from the database, never from the live stream, because “the sensor went quiet” is the absence of messages and a stream cannot see an absence. The language model reads the database, answers, and issues commands through the broker. If the model's provider is down, alerts still fire.
Never on your control system.
Watchman's devices talk to one place, the broker over TLS, and to nothing else. They never connect to a PLC, a PC or a control network. They need one thing from your network, an outbound connection on port 8883, so the guest Wi-Fi is enough.
One protocol, one broker.
Every device speaks MQTT over TLS to one broker, with its own credential and an access list that lets it publish only under its own topic and subscribe only to its own commands. A compromised camera cannot read the microphone. No HTTP from devices, no custom ports, no inbound connection to your site.
Bring your own sensors, through the same door.
Shelly and Tasmota devices, Zigbee sensors through Zigbee2MQTT, and LoRaWAN sensors through their network server already speak MQTT. An adapter translates their messages into Watchman's contract, under the same per-device credential and topic-scoped access list as our own nodes. Sensors that speak something else get a gateway, or a whole network, built for the site, publishing into the same contract. The contract does not change, so the rules, the baselines and the assistant treat them exactly like ours.
One database.
Registry, readings, events, commands, alerts and chat history in one Postgres. Nothing exotic to operate.
Alerts that behave.
A dead sensor is unknown, not still breaching, so it cannot alarm for ever. Thresholds carry a dwell time, so a half-degree sensor cannot flap across a limit. One message per breach with a cooldown, and a recovery message when it ends. “Went quiet” is a first-class alert, because it is the most common failure in the field.
Time is evidence.
Devices timestamp at source after clock sync. A device with a bad clock is stamped at receipt and flagged, never silently trusted. Every publish is acknowledged, so a device writing into a dead connection notices within a few intervals and reconnects. The cloud can see that failure and the device cannot, so only the device can fix it.
Accuracy is declared per metric, and the model quotes it.
Physics before statistics. Planned
A cold room is a heat leak and a cooling machine. The rate it warms while the compressor rests measures the leak; the rate it cools while the compressor runs measures the machine. Both are fitted daily from one thermometer and the run state.
Sound as numbers.
A-weighted level, peak, eight octave bands, crest factor, running fraction per interval, and speed by harmonic peak search. The 1 to 4 kHz bands are the ones that generalise to a real machine; low bands are wind on the microphone ports unless the array's coherence says otherwise, and it says so. An anomaly score against a per-machine baseline is published as triage, never as an alarm.
Sight as events.
The camera runs its own detector on its own chip and publishes counts and line crossings, plus a still on request or on event. The model sees one frame when asked and describes it.
Every front end is the same tools.
WhatsApp, the voice line, Claude Desktop over MCP and the HTTP API share one tool definition. What the model can do on one, it can do on all.
| Device link | MQTT 3.1.1 over TLS 1.3 · per-device credentials · topic-scoped access list · last-will presence |
|---|---|
| Network | Outbound TLS on port 8883 only · guest Wi-Fi is enough · nothing inbound · nothing on your control network |
| Sensor nodes | ESP32-S3 · DS18B20 ±0.5 °C · interval 10 s to 1 h · buffers readings offline and flushes with original timestamps |
| Microphone | 4-element USB array on Raspberry Pi · 16 kHz · features only · clips on event, off by default |
| Camera | On-device YOLO on an NPU · counts, line crossings, JPEG stills on request · no video |
| Your sensors | Shelly · Tasmota · Zigbee through Zigbee2MQTT · LoRaWAN network servers · anything that speaks MQTT, through an adapter · anything else, through a gateway or network we build for the site |
| Cloud | One Python service, one Postgres · rules every 10 s from the database · 30-day raw retention · stills and clips capped by age and size |
| Channels | WhatsApp for questions (Hebrew and English, typed or voice note) · Telegram for alerts (groups, acknowledge) · voice line · MCP for Claude · HTTP API |
| Model | Provider-agnostic · never on the alert path · budgeted per session and per day |
| Hosting | Single VPS today · dedicated instance per customer on request · fully on premises for the most demanding sites |
The pilot: one site, as many sensors as it needs.
No connection to anything you run, and no integration project on your side. We connect the sensors you already have, add ours only where you are blind, and you judge it on your own cold room.
What you get
- Monitoring for the assets you pick, on your sensors, ours, or both
- Baselines learned from your own first week
- A WhatsApp number your shift can ask, in Hebrew or English
- Alerts to the phones or group you name, with recovery messages
- A findings report at the end, including anything the first weeks of data say about your equipment
After the pilot:
Standard
Monitoring, alerts, the WhatsApp assistant, and a dashboard for the supervisor.
Full-Care
Everything in Standard, plus hardware care, replacement sensors, an annual visit and a baseline retrain.
Why “your AI watchman”?
Every plant had one. The night watchman walked the floor with a torch at three in the morning, listened at the compressor room door, put a hand on the cold room, and phoned you only when something was wrong. He had no dashboard. He had ears, eyes, and a sense of what the place normally sounded like.
That job is gone, and the plants that still need one cannot hire for it. Watchman is our answer: not another dashboard, but something awake at three in the morning that knows what normal looks and sounds like, and calls you when it changes.
Built by Segev Technologies
Watchman is a GoMachines.AI product, engineered by Segev Technologies, a Tel-Aviv R&D firm that has spent two decades building electronics, firmware, motion control and AI for robots and precision systems, for clients including Israel Aerospace Industries, Applied Materials, Elbit, Siemens, Mitsubishi Electric and HP.
We are machine builders. We didn't learn factories from a dataset.
Straight answers.
I already have sensors. Do I have to replace them?
No. Whatever your sensors are and however they talk, we connect them. Ones that already speak MQTT, like Shelly, Tasmota, Zigbee2MQTT and LoRaWAN network servers, take the short path. For the rest, we build the gateway or the network the site needs. We add our own sensors only where you have none.
We have a PLC. Do we need to connect it?
No, and Watchman never will. It watches from the outside, the way a person on rounds does. Nothing on your control network, nothing in the cabinet, no software on your computers.
Does it listen to conversations?
It cannot store them. The device sends a few numbers a second, not audio. A short clip is stored only on an event, only if you turned that device's clip flag on, and it expires. See “What it stores”.
Can all of it stay inside our plant?
Yes, for the sites that need it. The broker, the database, the rules and the model all run on a server on your premises, on your network. The checks in “What it stores” work the same way there.
Is it a safety system?
No. Watchman notifies. It does not stop machines and it is not a substitute for an interlock, a fire panel or an alarm system. If you need protection, buy protection. Watchman tells you when to go and look.
Does it record video?
No. The camera runs its own detector and sends counts, events and a still when asked. No video is stored or transmitted.
What if the internet drops?
The device keeps measuring and keeps its readings, then sends them with their original timestamps when it reconnects. The cloud notices the silence and tells you the device went quiet, which is itself the most common failure in the field.
Can it control anything?
Not unless you say so, twice. Actuation is off by default, per device and per person. Where you enable it, the assistant asks you to confirm in the same conversation before it sends the command. Reading never needs that.
Do the sensors talk to each other?
Through the database, yes. Every reading lands in one place, and the rules and the assistant reason over all of them together: an open door explains a warm room, a busy compressor with a cold room points at the seal. No sensor talks to another directly, so a compromised device cannot read its neighbours, and nothing is decided by a conversation you cannot audit.
Hebrew?
Yes, and English, typed or spoken. Every reading comes with its unit and how old it is.
Your plant is already telling you.
At three in the morning, to an empty room.
Let's put a watchman on the five things you worry about and see what they say.
Request a pilotWatchman · a GoMachines.AI product · engineered by Segev Technologies · Tel Aviv · segevtech.com