Telegram bot
A photo and a clip that plays right in the chat: class, camera, zone, time. Three languages. Clients see only operator-confirmed events.
Your cameras already see everything. Argus makes sure someone notices.
We connect to the RTSP cameras you already have. No new hardware, no third-party cloud.
2.0false alarms per 100 threat-free frames
92%of real threats detected
~200cameras — designed capacity
60 soperator response deadline
About ARGUS
ARGUS builds AI solutions for analysing and managing physical processes in the real world.
We analyse how people behave, how equipment runs and how vehicles move: the system understands events in a monitored area and turns data into concrete action.
Where it works
- Preventing offences and dangerous situations
- Detecting violations as they happen
- Finding events and people in large volumes of video
- Safety of people and work areas
- Control of equipment, vehicles and processes
- Detecting abnormal situations and risks
Our mission is to make the physical world observable, analysable and manageable with artificial intelligence.
Our first product is Argus 2.0: an AI monitoring system for live video, built on our own detection model, Neo.
Video surveillance answers what happened yesterday. Not what is happening now.
A person at a monitor loses attention after 20 minutes.
With 50 cameras you cannot watch every window at once.
Incidents get noticed after they are over.
Argus flips the logic: the system finds the event and brings a person to it.
Nine event classes
Neo detects four classes today. Five more come with upcoming updates.
CAM 03 14:07:51
knife 0.87KnifeThreatsliveCAM 11 22:41:03
gun 0.91FirearmThreatsliveCAM 07 03:14:22
fall 0.93Person fallingBehaviourliveCAM 09 23:18:12
fight 0.78FightBehaviourcoming soonCAM 05 01:52:40
person 0.71Presence in a restricted zoneBehaviourliveCAM 14 04:03:57
fire 0.84FireFire safetycoming soonCAM 14 04:05:11
smoke 0.76SmokeFire safetycoming soonCAM 02 09:26:33
no_helmet 0.81Missing hard hatWorkplace safetycoming soonCAM 02 09:27:08
no_vest 0.79Missing safety vestWorkplace safetycoming soonPPE violations only count where a hard hat and vest are mandatory.
What the system does not do
- no face recognition, no biometric records
- no licence plate reading
- no abandoned-object detection
Three steps from camera to person
-
We connect your cameras.
Any IP camera or recorder with an RTSP stream. Nothing changes in your existing setup.
-
Neo watches the frames.
The Neo model runs on a GPU computer at your site. One node handles dozens of cameras.
-
The event reaches a person.
A snapshot with a box, a short clip, camera, zone, time. In Telegram, in the web console, on the operator's desk.
Your camerasRTSP
GPU nodeon site · Neo model
events only
Central serverprivate network
OperatorsTelegramClient consoleEmail reports
A system that shouts “weapon!” forty times a shift gets switched off within a week.
Typical neural networks mistake a phone for a pistol and headlights for fire. Neo is built for street cameras: real angles, poor lighting, long range.
17×fewer false alarms with Neo than with a typical off-the-shelf model
| Typical off-the-shelf model | Neo | |
|---|---|---|
| False alarms per 100 threat-free frames | 33.8 | 2.0 |
| Share of real threats detected | 78% | 92% |
| Average confidence on correct detections | 0.72 | 0.81 |
| Processing speed, frames/s | 18.6 | 27.8 |
| Stable detection run on real video | 0.8 s | 10 s |
Measured on an internal control set. Results on a site depend on angles and camera quality.
Seven layers of protection against false alarms
K-of-N frame confirmationAn event is created only when the object is seen consistently.
Per-class thresholdsThe bar for a weapon is higher than for a person.
Detection zonesReact to the gate, ignore the road behind the fence.
Per-camera cooldownOne incident, one message.
Object trackingThe event is tied to a person, not a frame.
Night threshold profileAn IR image is a different visual reality.
Site adaptationNeo learns from frames of your cameras.
Neo gets more accurate with every week of operation.
Neo is not a frozen out-of-the-box model. Operators give one-click verdicts, Neo takes your cameras and lighting into account — and works better than on day one.
weapon 0.41typical off-the-shelf model — 0.8-second flashesweapon 0.86Neo — a stable run of up to 10 secondsFour channels, one source of truth
Operator console
Event feed, live camera view on click, statistics with a heatmap, zone editor, audit log. Works on a phone.
Client console
A read-only console: charts and a feed of confirmed events for the client's own cameras.
Email reports
Weekly or monthly: charts, a summary by class and snapshots of key events.
Operators and SLA
- Events go to the least loaded available operator.
- Review deadline is 60 seconds, configurable.
- One-button verdict, including straight from Telegram.
- Overdue events are not lost — they land in “Unreviewed”.
Your video stays with you
Video never leaves the siteAnalysis on your hardware; only events go out.
No third-party cloudsAll neural networks run locally.
Private networkNodes connect themselves; no inbound ports.
No biometricsFaces are not recognised.
Access separationEveryone sees only their own cameras and sites.
AuditEvery action is logged: who, what and when.
What's next
We only promise what already works. Here is what is in development.
- Fight — recognition from the motion of several people.
- Fire and smoke — training the fire-safety model.
- Hard hat and vest — training the workplace-safety model.
- Site map and floor plans in the operator console.
- Day/night guard schedules and camera tampering detection.
From audit to full deployment
Audit, 1–3 days
We review cameras, angles and lighting, and pick the classes that fit.
Pilot, 2–4 weeks
A node, 5–20 cameras, zones and thresholds. Real events, seen first-hand.
Tuning for the site
We collect operator verdicts and adapt Neo to your cameras.
Full deployment
All cameras, operators, Telegram for clients, reports for management.
Start with 10 cameras for two weeks. If false alarms outnumber useful events, you lose nothing.
Request a pilotWhat people ask before a pilot
We already have video surveillance
Argus connects to your cameras alongside the archive and changes nothing in it. We do not replace surveillance; we add attention to it.
What if the network goes down
Analysis runs locally and does not stop. Undelivered events are queued and sent once the connection is back.
Neural networks make mistakes all the time
That is why every event passes seven layers of filtering. The result: 2.0 false alarms per 100 frames versus 33.8 for a typical model.
Our video will end up in the wrong hands
Video never leaves the site: only a snapshot and a short clip go out. No third-party clouds, no inbound ports.
Do we need to train staff
Telegram is enough for an operator: the event arrives on its own, the verdict is one button. The web console takes one shift to learn.
Does it work at night
Yes. Neo works on IR imagery with a separate night threshold profile — the night shift is exactly the time this system exists for.
See the events from your own cameras
Leave a request — we reply within one business day.
Talk to us directly
Faster than the form — message us directly.
Contact details are added at publication (file js/config.js).