Our mission.
Security operations run on cameras, and the number of cameras keeps growing. The number of people watching them does not. Somewhere tonight a guard is looking at a wall of quiet tiles while the one that matters sits in the corner of a cycling layout. That is the problem we exist to solve.
Our mission is simple: help every guard cover more cameras with clearer priorities, while keeping human judgment at the center of every decision. The AI surfaces, explains and organizes. The guard decides, dispatches and answers for the outcome. We think that is the right division of labor for security, and we build for it deliberately. Read why in human-in-the-loop security AI.
Our approach: build around the guard’s screen.
Most video analytics start from the camera or the VMS. That means integrations, credentials, compatible hardware and a project for every site. The operators who need help most, small and mid-size guard companies and monitoring centers covering a mix of client systems, are the ones least able to pay for that.
So we started somewhere else. Semaphor works from the monitoring screen guards already watch. It captures the rendered display, separates the camera panes, follows each view as layouts change and marks any camera that leaves the screen as unobserved. No camera replacement, no new VMS, no vendor API. See works with existing camera software and how it works.
Our engineering.
We build on real video and we check our work on real video. Our detection, tracking and scene-report pipeline is developed against recorded footage, including public research datasets like the one in our recorded demo, and evaluated on held-out clips rather than the footage it was tuned on. We keep camera identity, tracking and AI interpretation as separate, testable steps so we can see exactly where something went wrong.
Evidence comes first. Every alert keeps the original frames beside the AI’s assessment, with exact SHA-256 hashes and perceptual hashes, so an assessment can always be checked against the pixels it came from. Alerts publish before the AI finishes, so guards never wait on a model. And we track people and vehicles per camera without facial recognition. See evidence and audit trail and security and privacy.
Where we are going.
We are actively building the next layer: spatial scene models that place detections on a floor plan, object memory that persists across restarts, reading camera labels and named landmarks, event timelines for fast review, and fire and smoke detection trained and evaluated offline and designed to complement certified fire alarm systems. Pilot partners get early access. See what’s next.
Work with us.
The best way to understand Semaphor is to see it on your own monitoring screens. Start a pilot on one station, or book a live demo.