Built on a working loop.
Everything on this page extends a loop that already runs: screen capture, camera view tracking, person and vehicle detection, alert-first review, AI scene reports and a guard inbox. If you are new here, start with what Semaphor is. What follows is active development — trained models, evaluations on real video, protocols with versioned contracts. Pilot partners get early access and shape what ships first.
Fire and smoke detection.
Fire and smoke detection is in active development, with early access for pilot partners. Our team has trained an RF-DETR-based smoke and flame detector and evaluated it offline on held-out images and real video replays, including dashboard-sized, compressed panes that mimic what a monitoring screen really shows. On top of the detector we are building temporal checks that look for smoke developing over time instead of trusting a single frame, plus inexpensive regional hash and texture descriptors that help separate haze, steam and glare from the real thing.
The goal is simple: when smoke starts rolling across a loading dock camera at 3 a.m., the guard should see it on a card, with the frames, before anyone has to notice it on a wall of tiles. It is designed to complement certified fire alarm systems, not stand in for them — life-safety detection stays with your listed alarm panel. Read the background in our guide to AI fire and smoke detection on cameras.
A spatial scene model.
Cameras show pixels. Guards think in places: the north gate, the second loading door, the path behind building C. We are building a spatial model that projects person observations from a camera onto a calibrated floor plan, with top-down, side and rotatable views. Every placed position links back to the original frame and box it came from, so a map dot is always one click from its evidence. The first version runs on real research video with supplied calibration; the direction is full scene graphs that let the AI reason about where things are, not just what they are.
Object memory.
Today, track IDs are clip-local: numbered hypotheses inside one camera view. Object memory keeps people and vehicle records — with their local IDs, observation history and evidence-linked appearance notes like clothing color or vehicle type — across restarts. A guard asking “is that the same white van from an hour ago?” gets a history, not a guess. Appearance vocabulary is deliberately limited: no race, ethnicity, gender or age inference, and no facial recognition. That is a privacy line we drew on purpose; see security and privacy.
OCR camera labels and landmark maps.
Most video walls print a camera name on every tile. Semaphor’s local OCR is learning to read those labels, confirmed across several samples before it trusts them, so an alert can say “Dock East 03” instead of “view 7.” Alongside labels, the vision model proposes named fixed landmarks — poles, doorways, hydrants, fence lines — and Semaphor verifies they persist in the scene before placing them on a camera’s map. The result is stronger feed identity when layouts change and plain-language locations in every scene report.
Severity prioritization.
The inbox already sorts Critical, High, Unassessed and Low, with guard overrides. The next step is automated severity scoring that operators can trust: site-specific rubrics, independently labelled dangerous and harmless events, held-out cameras and per-level scoring, so a Critical badge means what your post orders say it means. See alert prioritization for how the queue works today.
Symbolic event timelines.
We call them flip books. Each candidate event becomes an ordered, compact text timeline — P01 moved right, V02 stopped, region M01 changed — with every symbol pinned to its source image and hash. Guards scan a timeline in seconds instead of scrubbing footage, and the AI reviews the same ordered record alongside the original frames, which makes its reasoning easier to check. Flip books already run on real video in our evaluations.
Second-opinion model comparisons.
We run candidate models side by side on the same event records and typed questions — uncertainty handling, policy exceptions, hostile text on screen, paraphrased and repeated inputs — without touching live alerts. Only models that hold up earn a place in the guard loop.
How we build.
- Real video, held out. Models are trained on one set and evaluated on footage they have never seen, including compressed, dashboard-sized panes.
- Evidence first. Every new layer links back to original frames and hashes.
- Guard in control. New capabilities add context to the card. The guard still decides.
- Shipped when proven. Features move into the live guard loop once they pass evaluation on real sites.
Get early access.
Pilot partners get roadmap features first, on their own screens, and help decide what ships next. Start with one monitoring station through our pilot program, or book a live demo and tell us which of these your guards need most. The loop that runs today is the foundation it all ships on.