Visionman Every camera, watched.
Visionman watches your cameras the way a person at the monitors would: one picture every second, all day, on every camera at once. It writes down what it sees, answers questions about any moment, and sends you a message with a photo when an event you described happens. You set all of it up in plain words.
It describes what it sees, never who someone is. Your cameras stay off the internet, and pictures are deleted on a schedule you set.
Works with our own camera, which runs an object detector on its own chip, and with the IP cameras you already have, through a small gateway on site.
- 14:31:50Two people at the entrance. One carries a red bag.
- 14:32:00The red bag is set down beside the bench.
- 14:32:10Both walk out toward the street. The bag stays by the bench.Rule 2 · yes · frames 3, 7
- 14:32:20No one near the bag.Rule 2 · yes · frames 1, 9
Rule 2: Tell me if a bag is left alone by the entrance.
A red bag beside the bench, no one near it since 14:32:10.
Your cameras record everything. Nobody is watching.
Most camera footage is watched once: the day after, by someone looking for the moment they missed.
A site with twelve cameras records close to 300 hours of video a day. A person at the monitors can follow a few screens for a few hours, and at three in the morning nobody is looking at all. So cameras mostly work after the fact: something goes wrong, and someone scrolls through hours of recordings to find when it started.
Nobody at the monitors.
Watching screens for hours is hard, and attention fades long before the shift ends. Most sites cannot staff the monitors around the clock, so most cameras are never watched live.
Motion alerts that cry wolf.
Motion detection fires on shadows, rain, headlights and the neighbour's cat. After a week of false alarms, nobody opens them.
An hour to find one minute.
When something does happen, finding it means scrubbing through recordings, camera by camera. A simple question, such as when the bag first appeared, takes an hour to answer.
It watches the way a person would. On every camera, all day.
Connect your cameras.
Use our camera, which counts and tracks objects on its own chip, or connect the IP cameras you already have through a small gateway on site. Each camera sends one picture a second over an encrypted connection. All it needs from your network is one outbound connection, and your cameras stay off the internet.
It writes down what it sees.
Every ten seconds, Visionman reviews the last ten pictures and writes down what happened: where people were and what they were doing, every object in view, and anything it could not make out. Tell it what each camera covers and what matters there, and it watches with that in mind. The result is a written record of the whole day that you can search.
Ask it, or let it call you.
Ask about any moment in a chat message, in your own words, and get an answer with the photo behind it. Describe an event in one sentence, and Visionman checks for it every ten seconds and messages you, with a photo, when it happens.
Every answer points to the pictures it is based on. When the pictures cannot settle a question, because it is too dark, too far or hidden, Visionman says so instead of guessing.
Ask about any moment. In plain words.
Questions go to a chat, in whatever language you think in. Visionman answers from its written record, and when a question needs a closer look, it goes back to the stored pictures around that moment at full resolution.
Each answer says how sure it is, which pictures it used, and how much of the period it actually saw. An hour when a camera was offline is reported as an hour it did not see, never as an hour when nothing happened.
Tell it what matters. It tells you when it happens.
A rule is one sentence about an event or a situation, written the way you would say it. Visionman turns it into a question, asks it of every camera the rule covers every ten seconds, and messages you when the answer is yes.
“Alert me if someone climbs over the fence.”
“Tell me if anyone goes into the server room.”
“Tell me if a car comes in through the exit.”
“Message me when 4 or more people are running.”
“Tell me if a bag is left alone by the entrance.”
“Let me know if the loading bay door stays open.”
“Tell the shift group if there is smoke in the storeroom.”
“Tell me if anyone is in the shop after closing.”
“Let me know when a vehicle stops in front of the gate.”
What arrives on your phone
The photo that shows it, the rule in your words, the time, and a line on what was seen. When the situation ends, a second message says so and how long it lasted.
“Tell me if a bag is left alone by the entrance.”
A red bag beside the bench, no one near it since 14:32:10.
14:38:40 · Over after 6 minutes.
Checked before it calls you
- One-off events, such as someone climbing the fence, alert on the first window that shows them. Lasting situations, such as a door left open, must first hold in 2 of 3 ten-second windows. Both are adjustable per rule.
- Each window is judged on its own pictures, so one misread picture cannot repeat itself.
- An optional closer look at full resolution before the message goes out.
- Ignore zones for screens, posters, mirrors and glass that show people who are not there.
- Quiet hours, a pause between alerts, and one message per event, not one per minute.
- Each rule goes to the people or the group you choose.
Rules are made in the same chat, by the people you allow to make them.
It also tells you when it cannot see.
On a recording, a covered or disconnected camera looks just like a quiet night. Visionman checks every picture it receives and tells you as soon as the view or the connection fails.
| What happens | What you get |
|---|---|
| The camera starts | “Started at 07:02, view OK”, with a photo so you can check the aim |
| The connection drops | “Lost its connection at 14:31” |
| Connected, but no pictures arrive | “Connected but sending no pictures since 14:29” |
| The view is covered, blocked or dark | “The view is dark or blocked”, with the last picture |
| Pictures stop being reviewed | “Not watching”, and the reason |
| Any of these clears | One line saying so, and how long it lasted |
A message that does not get through is sent again until it does, never twice, and a recovery never arrives before its alert.
What leaves the camera, what is stored, and for how long.
Video analysis in a workplace is a serious thing. Here is exactly what Visionman keeps, and what it does not.
Full video stays with you.
Recording stays on the camera or on your own recorder, as it does today. The camera runs its object detector on its own chip and sends one picture a second. Seconds in which nothing changed are not sent at all.
Kept as evidence, then deleted.
Pictures are kept for the period you set, from an hour to a week, so questions and alerts can show what happened. Then they are deleted automatically. Backups never include pictures.
Text you can search, for a set time.
What Visionman wrote down is kept for the period you set, 90 days by default. Only the people you register can search it, each limited to their own site's cameras, and every question is logged: who asked, about which camera, and when.
People are described, never identified.
Visionman can note what someone wore and carried, because that is what makes “a man with a red bag” answerable. It never determines who anyone is and does no face recognition. Where descriptions are not wanted, switch them off for that camera, and erase what was already stored.
The analysis runs on paid API tiers that do not use your pictures to train models.
Works alongside your security systems. Visionman is a monitoring and search tool: it tells your team when to look and helps them find what happened. Alarms, access control and guards keep doing their jobs, and Visionman adds an attentive pair of eyes on every camera.
Three checks your own team can run
- V1 · Read the record.Everything Visionman wrote about a camera is a row in its database. Open it.
- V2 · Look for an old picture.Search for a picture older than its retention period. There is none: deletion runs every minute.
- V3 · Flip the switch.Turn off people descriptions for a camera. From the next ten seconds on, its record holds positions and activity only.
Works with the cameras you have.
Visionman cares about what a camera sees, not who made it. Use ours, connect yours, or mix both on one site.
A camera that counts on its own.
A compact camera with an object detector on its own chip. It recognises 80 everyday kinds of object (people, animals, vehicles, bags, furniture and more) several times a second and marks the ones moving fast. It records video locally, and on request it sends a closer look, at 15 frames a second, around a recent moment.
The cameras already on your walls.
Existing IP cameras connect through a small gateway on site, a mini PC that reads each camera's stream, runs the same detector, and sends the same pictures and counts. Where your recorder already records, it keeps doing so; otherwise the gateway records each stream itself.
Either way, cameras and recorders stay inside your network, and only one encrypted outbound connection leaves the site.
Placement matters more than resolution: a camera that sees the entrance clearly answers questions about the entrance well.
Under the hood.
This is what Visionman does, in the terms your engineers and your security team will check.
One picture a second, reviewed ten seconds at a time.
Each camera sends one JPEG a second. Visionman groups them into 10-second windows on capture time, waits a few seconds for late pictures, and has a vision model review each window. If reviews fall behind, windows are merged until they catch up, so a slow minute never builds a queue.
Rules are questions that need evidence.
Every active rule becomes a yes or no question asked in every window. A yes counts only when it cites the frames that show it. A rule about a one-off event fires on the first window that shows it; a rule about a lasting situation fires when the yes holds in 2 of 3 consecutive windows (both adjustable). Rule answers come from each window's own frames: the running story of the scene feeds the notes, never the evidence.
“Cannot determine” is an answer.
Too dark, too far, hidden or blurred gives “cannot determine”, which neither fires a rule nor clears one. The record shows which windows were reviewed, skipped as unchanged, or missing.
A second opinion on the camera.
The on-board detector counts 80 object types several times a second and marks fast movers, at no processing cost. Its counts are stored with every window, as a second index for search.
Unchanged seconds are skipped.
The camera compares each picture with the last one it sent, block by block, and skips seconds in which nothing changed. It still sends a picture at least every 30 seconds, which also proves it is alive.
A closer look when it matters.
A question about a specific moment goes back to the stored pictures around it, 30 seconds by default, at full resolution. The answer comes with the picture it rests on.
A memory you can search.
The written record is searchable by any word, summarised by the hour and by the day, and grouped into episodes, so “how many times” has an answer. For a question the record cannot answer, Visionman can re-read the stored pictures of a whole period in the background and message you when it is done.
Each camera walled off.
Cameras speak MQTT over TLS to Visionman's own broker, each with its own credential and an access list that lets it publish only under its own topic. The service itself listens on no inbound port.
Time is evidence.
Pictures are stamped by the camera when its clock is synchronised, and by the server, flagged, when it is not. Questions and messages use the site's local time.
Built to keep going.
Messages are retried until delivered and never sent twice. Rule and health state live in the database, so a restart picks up where it stopped, and the service reconnects on its own after a database or network restart.
The server decides who sees what.
Each user belongs to a site and has a role: viewer, operator or admin. The server, never the model, decides which cameras a question may touch, and every question is logged with its outcome.
| Input | 1 JPEG per second per camera, up to 1280×720 · unchanged seconds skipped, with a picture at least every 30 s |
|---|---|
| Review | 10-second windows, 5 to 60 s per camera · vision model at low detail for the record, full resolution for questions and confirmations |
| On-camera detector | 80 everyday object types · 4 or more runs a second beside 15 fps local recording · per-type counts and fast movers |
| Device link | MQTT 3.1.1 over TLS · QoS 1 · per-camera credentials · topic-scoped access list · last-will presence |
| IP cameras | RTSP through an on-site gateway · sub-stream or main stream per camera · records by stream copy, or leaves it to your recorder |
| Rules | plain-language text, for events and lasting situations · events on the first window that shows them, situations on 2 of 3 windows, adjustable · recovery after 3 clear windows · pause between alerts · quiet hours · ignore zones · optional confirmation look |
| Questions | any moment within picture retention · search across the written record · hourly and daily summaries · episode counts with coverage |
| Channels | Telegram and WhatsApp · recipients and groups per rule · a photo with every alert |
| Health | start with photo · lost connection · no pictures · dark or blocked view · not watching · each recovery with its duration |
| Retention | pictures 1 hour to 7 days · written record 90 days by default · per-camera purge on request |
| Access | registered users per site · viewer, operator and admin roles · every question logged |
| Server | one Python service, one PostgreSQL database · the only open port is the broker's TLS port · nightly backups that never include pictures |
Start with one site.
We connect your cameras, write the first rules with you, and tune them on your own site. You judge Visionman on your own cameras, in a normal week.
What you get
- Visionman on the cameras you pick: ours, yours, or both
- A description of each camera's view and what matters there, written with you
- Your first rules, set up and tuned with your team
- Questions and alerts in the chat app your team already uses
- Camera health messages from the first day
- A review after the first weeks: what it caught, what it flagged, and what to adjust
Tell us about your site
How many cameras, what they cover, and what you would like to know. We will come back with a plan.
Why “Visionman”?
Every control room has one: the person at the monitors, watching a wall of screens, noting what happens, and picking up the phone when something looks wrong. It is one of the hardest jobs in security. Nobody can watch twenty screens for eight hours, and nobody remembers what screen nine showed at 14:32 last Tuesday.
Visionman does that job on every camera at once, all day, and writes it all down. Your people decide what matters, and hear about it when it happens.
Built by Segev Technologies
Visionman 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 computer vision for robots and precision systems, for clients including Israel Aerospace Industries, Rafael, Elbit, Applied Materials, Mitsubishi Electric and HP.
Computer vision is part of our daily work, from lab instruments to robots on the factory floor.
Straight answers.
Does it recognise faces or identify people?
No. It describes what is visible, such as clothing, colours and what someone carries, and never who someone is. There is no face recognition, and the descriptions themselves can be switched off per camera.
Do we need new cameras?
No. Your IP cameras connect through a small gateway on site. Our own camera adds an on-board object detector and local recording, and one site can mix both.
Does video leave our site?
Full video stays on site. One picture a second goes out over an encrypted connection for review, is kept for the period you set, and is then deleted automatically. See “What it keeps”.
How fast do alerts arrive?
A one-off event, such as someone climbing the fence, is usually reported within about 20 seconds. A lasting situation, such as a door left open, first has to hold in two of three ten-second windows, so it arrives within about half a minute. That wait is what keeps a single odd picture from waking anyone up.
Can it catch something that happens in a split second?
It sees one picture a second, so it is built for situations that last a second or more: people running, a crowd forming, a bag left behind, a door left open. For anything faster, the camera's own recording keeps the full video.
Does it work at night?
It sees what the camera sees. Cameras with infrared give it a night view in grey. Where the view goes dark, it tells you the view is dark instead of reporting a quiet night.
Who can ask questions?
Only the people you register, each limited to their own site's cameras and to the role you give them. Every question is logged.
What if it gets something wrong?
Like a person at the monitors, it will sometimes miss something or misread a picture. That is why every answer cites its pictures, every alert carries a photo, a lasting situation must hold over time before it alerts, and an optional closer look can confirm before anyone is called.
What happens if the internet drops?
The camera keeps recording locally. Visionman notices the silence, tells you the camera lost its connection, and tells you again when it is back.
Which languages?
Ask in any language your team uses, and write rules in it too.
Your cameras see everything.
Now you can ask them.
Let's put Visionman on the cameras that matter most to you and see what it tells you.
Request a demoVisionman · a GoMachines.AI product · engineered by Segev Technologies · Tel Aviv · segevtech.com