Public-safety AI · Privacy-first
Every camera watched. Every second counted.
AlmanoEye turns a city's existing CCTV network into a real-time safety instrument. It detects violence, medical emergencies and dangerous crowd conditions as they begin, and puts a reviewed, evidence-backed alert in front of an operator in seconds rather than minutes.
Working pilot · live demonstration available
- ≈10 s
- from the first seconds of an incident to a verified operator alert, measured end to end
- 91%
- detection of fallen, motionless persons, up from 73% with standard models
- 68%
- of benign candidate events dismissed by deep review before they ever reach an operator
- 0
- faces recognised, identities stored or frames sent to any cloud, by architecture rather than by policy
The system, running
Two minutes inside the operator's screen.
A walkthrough of the working pilot: live detection on a street camera, a single incident opened and reviewed frame by frame with the reasoning shown, and the deployment view a commander would use to place units.
The problem
Cameras record. Almost no one watches.
European cities have invested heavily in CCTV, but the dominant use is forensic: footage is retrieved after the fact, to investigate what has already happened. The research on control rooms is consistent — one operator can genuinely attend to only a handful of screens, and vigilance decays sharply within the first half hour of a shift.
The gap between recording and watching is measured in minutes, and in an emergency minutes are the currency of survival. A person who collapses in view of an unmonitored camera has not been helped by that camera.
How it works
A two-stage system that reviews itself before it alarms anyone.
AlmanoEye mirrors how a good control room works — a fast watcher and a careful reviewer — and applies that pattern to every camera, every second.
-
Stage one
AlmanoEye Light
- Watches every frame of every stream in real time
- Tracks people as anonymous skeletons, never as identities
- A learned motion model flags fight-like interaction, falls, panic and crowd pressure
continuous · 25 fps per camera
-
Stage two
AlmanoEye Heavy
- Replays the six seconds around each candidate event
- Detailed hand analysis: grabs, fist postures, contact between silhouettes
- Confirms the incident, or dismisses it with a written reason
≈7 s per episode
-
Always in command
The operator
- Receives only reviewed incidents, each with a clip, a skeletal replay and the full evidence log
- Confirms or dismisses in one click
- Every decision becomes training data, so the system learns the city's own patterns
a human decides · every time
What it watches for
Three kinds of emergency, caught as they begin.
- Violence, as it starts
- A compact model reads limb kinematics rather than pixels, so it recognises the signature of an assault within the first seconds — in daylight, at night, in rain, in a crowd.
- Medical emergencies
- Someone who falls and stays down triggers an escalating medical alert, including when the person becomes hard to see. A dedicated model is fine-tuned specifically on people lying on the ground.
- Crowd crush, before it is lethal
- Continuous crowd-pressure estimation follows the physics of crowd disasters and warns while compression is still building — relevant for festivals, stations and match days.
Privacy by architecture
Safety without surveillance of identity.
AlmanoEye analyses how bodies move, never who people are. Detection runs on anonymous skeletal geometry: the system contains no facial recognition and no biometric identification. Everything runs on-premise, and no video frame leaves the building.
This is a property of the architecture, not a setting that could be switched back on.
- No facial recognition
- No biometric identity
- Runs on-premise
- Skeleton-based analysis
See it running on your own cameras.
AlmanoEye is a working pilot and we demonstrate it live. Tell us what you are responsible for keeping safe, and we will show you what the system sees.
Book a demonstration