Product
How Aegis Vision works
Real-time AI detection of weapons and safety incidents on a site's existing cameras, processed on the site's own hardware.
The detection cascade
Detect, verify, alert, respond. A fast detector scans every frame; a larger model verifies the detection and the alert is raised to the site's guard console; two classifier models then grade the incident; a guard confirms, dismisses or escalates. The system is A multi-model computer-vision system of more than one billion parameters in total; the heavier models only run on frames the fast detector has already flagged, which is what keeps every camera covered in real time on a single server.
What it detects
Object detection
Handgun, rifle, knife or blade, explosive device, blunt melee weapon.
Body movement and context
Fights and assaults; a person who collapses and stays down; an unattended bag; a guard post left empty. Behaviour modules have no published accuracy figures yet; figures will be published once measured on real deployments.
What it does not do
Weapon and incident detection works on objects and body movement. It does not identify people, match faces against any database, or build profiles. The guard-post module recognises only the enrolled guard's face, only on the camera covering that post, and only when a site enables it. No other person is identified.
Alerts and escalation
On-site guard console (local web dashboard); SMS and voice call to designated phones through an on-site GSM modem; no consumer messaging apps. Every alert is reviewed by a person at the site before any external escalation; the system never contacts emergency services without a confirmed alert except on the highest-severity tier configured by the site.
Three independent checks before an alarm sounds; borderline detections must persist across frames and be confirmed by a second model, and the classifier stage can stand an alarm down while keeping the event logged.
Deployment
On-premises. Inference runs on a GPU server inside the building; full-resolution video never leaves the site. The guard console runs on the site's local network and keeps working if the internet connection drops. Works with existing IP cameras and recorders (RTSP streams); no camera replacement required.
Compute: One GPU server per site, specified at the site survey according to the number of cameras, supplied by Aegis Vision or by the school to that specification.. Network: The server reaches the camera streams on the school's local network; outbound traffic is limited to the encrypted private network used for support and updates.. Integrations: No integrations are published; connections to a school's VMS or incident-management tools are scoped per deployment..
Per-site tuning
Each deployment is fine-tuned on the site's own confirmed alerts; retraining and model updates are started by a human operator, never automatically. On one deployment, a detection the general model found in 1 of 10 test passes was found in 10 of 10 after a short fine-tune on that site's own footage, with no new false alarms.
Evaluation
Every pilot is measured on the school's own footage against criteria agreed in advance: time to alert, false alarms per camera per day, and guard feedback. Aggregate figures and methodology will be published on aegisvision.org once measured; nothing is published that would help someone evade detection.