Proactive security from any camera you fly, drive or wear.

AInspector uses visual GenAI to turn live and recorded video from drones, robots, vehicle and body-worn cameras into alerts your team can act on. It runs on an edge AI gateway at your site.

Talk to us See the use cases
3AI agents on every scene
2must agree before an alert
0faces or plates stored
Rail corridor inspection Person approaching Track bed, drawn areaCritical. Person on tracksStop trains on this line

Five use cases, ready to run.

Each one comes with its own rules and responses. Your team draws the areas to watch once, and the rules of each area apply.

Rail corridor

Know what is on the line before the next train.

For track patrols and rail control rooms. People, vehicles, animals, fallen objects, fire and vegetation on or near the track. Trains and rail plant are recorded without raising an alert.

  • Critical at onceA person on the track, or a vehicle in the corridor.
  • Every alert says what to doStop trains on this line, with a full-resolution clip of the moment.
Level crossing inspection Vehicle nearbyCritical. Person on tracksStop trains on this line
Perimeter

A fence line that reports what approaches it.

For compounds, depots, substations and data centres. A person inside the fence or approaching it is critical. Traffic on the road outside stays quiet.

  • Nearby is different from insideSomething near the fence alerts only if it comes closer or stays.
  • Send security to the placeThe alert carries the picture, the time and the response.
Perimeter security Vehicle nearbyCritical. Person inside the fenceSend security to the place
Border line

Know who is at the line, and on which side.

For border forces and coast watch. People, vehicles and boats that approach or cross the line raise an alert. A person in the water is routed as a rescue.

  • A group is one alertSeven people walking together raise a single alert.
  • Released by a personNo border alert leaves the site until a reviewer accepts it.
Boat on the waterBorder lineCritical. Group of 4 crossed the lineSecurity alert, held for a reviewer
Site patrol

Check the yard after hours, every night.

For depots, yards and substations, by drone, ground robot or patrol vehicle. After hours, a person or moving vehicle on site raises an alert and an open gate is flagged. The parked fleet stays quiet.

  • Hours from a trusted clockWorking hours are read from the flight log or the live feed, never from a file's date.
  • Police only on evidenceThe response sends a guard first, as remote monitoring rules require.
Depot after hours ParkingGate opened after hoursHigh. Vehicle on site after hoursSend a guard. Police on evidence.After hours, 02:14
Venue

Crowds, vehicles and exits at events.

For stadiums, festivals and official visits. A vehicle moving towards a crowd is critical at once. A blocked exit, a person on a roof or a flare outside its area each has its own priority.

  • People are countedA crowd is counted as points. Nobody is identified or stored.
  • Shelve with a reasonA planned activity can be put on hold for a set time. The alert returns on its own.
Venue forecourt watch Crowd area 1Exit route 1Effects areaRestricted roofVehicle nearbyHigh. Stopped in Exit route 1Security alert, exit blocked

In every use case, no faces or number plates are stored and nobody is identified. No lesson taught by an operator can silence an alert about a person or a vehicle.

Three AI agents check every scene.

A computer vision detector and two visual GenAI models look at each scene on their own. The GenAI models read the scene the way a person does and describe it in plain words. An alert needs two of the three to agree, which cuts false alarms without hiding real ones.

  • DetectorFinds objects and tracks them across every frame.
  • ObserverA visual GenAI model that describes the whole scene.
  • VerifierA second visual GenAI model that checks each candidate up close, without seeing the observer's answer.
  • AgreementOne agent alone sends a finding to review. Two or three raise the alert.
One sceneWhat reaches the operatorscroll sideways
A scene goes to the detector, the observer and the verifier; the verifier receives the detector's candidates and never sees the observer's answer; the result is a count of AI agents, and an alert only by the rules of the place Scene as the camera saw it Detector finds and tracks Observer describes the scene Verifier checks each candidate candidates kept separate Seen by three of three Alert by the place's rules

Rules that depend on the place.

Every finding is placed in the area you drew, near it, or elsewhere in view. Only the rules of that place apply.

  • In the areaCan raise an alert.
  • NearbyAlerts only if it comes closer or stays.
  • Elsewhere in viewKept for context, never an alert.
  • AnywhereFire, smoke and a person on a track always alert.
  • When the area cannot be foundEverything is treated as inside, and the operator is asked to draw it.
Elsewhere in viewNearby, approaching In the corridorAlert by the rules of the place Danger area

One workspace.

The picture, a timeline with a lane for each agent, and the finding sheet, on one screen. Day and night modes, and a wall mode for the control room.

The AInspector workspace at night: a gatehouse and fenced truck yard seen from a drone, a bus marked as a critical alert, the timeline with a lane per agent, and the finding sheet with the response to invoke the vehicle plan and inform police
A bus in a fenced truck yard. The detector, the observer and the verifier all reported it, so it is a critical alert with its response: invoke the vehicle plan and inform police.
A live inspection of walkers at a rural stile: the analysis panel shows what each agent sees frame by frame, and a doubtful finding is marked likely not there with no alert raised

Checked before it alerts

The analysis runs live, frame by frame. Here the observer thought a walker was lying on the ground. The verifier looked and did not confirm it, so no alert is raised and the finding is suggested for dismissal.

Settings, Memory: the lessons reviewers taught the installation, one waiting for a person to accept it, with its history and expiry date

What it has learnt

Every lesson your team teaches is listed with its place, its history and its expiry date. A new lesson changes nothing until someone accepts it.

Your team teaches it what is normal.

Mark a finding as not a hazard here, and the next patrol of that place is quieter. An administrator approves every lesson, and lessons expire.

  • Accept or dismissOne click, with five seconds to undo.
  • Not a hazard hereA lesson about one place. It needs two approvals before a repeat stops alerting.
  • People and vehicles always alertNo lesson can silence them.
  • Lessons expireAfter 30 days by default, 90 at most.
A lesson's lifetwelve seconds, loopedscroll sideways
Animated sequence: a reviewer marks a finding as not a hazard here, the lesson waits for a person, one administrator accepts it and a repeat then needs two sources, a second accepts it and a repeat is recorded and not alerted, and the lesson counts down to its expiry Small white building High, in the corridor, seen twice Accept Dismiss Not a hazard here Lesson for this place Not a hazard here, rail corridor waits for a person 1 of 2. A repeat needstwo sources to alert 2 of 2. A repeat isrecorded, not alerted Expires by itself 30 days by default, 90 at most People and vehicles are never quieted, however many people agree.

Evidence for everything. Your video stays on site.

Built for defence and rail. Every model call is recorded, and nothing leaves the gateway unless an administrator allows it.

  • A full audit trailEvery model call, alert, review and download is recorded, and the record cannot be edited.
  • Clips at full resolutionExport the moment around any finding.
  • Retention you setOld evidence is deleted on schedule, and each deletion is logged.
  • No cloud by defaultHosted models stay off unless an administrator turns them on.
  • Five rolesViewer, operator, reviewer, administrator and auditor, enforced by the server.
  • No route outThe AI agents run on a network with no connection to the internet.
Audit recordchain verified
Animated audit record: entries for model calls, alerts, reviews, downloads, lessons and retention scroll past, each linked to the one before 09:14:02Observer model call recorded 09:14:03Verifier model call recorded 09:14:05Critical alert raised, person on tracks 09:15:10Alert acknowledged by an operator 09:16:00Clip downloaded by a reviewer 09:20:41Lesson proposed, not a hazard here 09:21:03Lesson accepted by an administrator 23:59:00Old evidence deleted, tombstones written 09:14:02Observer model call recorded 09:14:03Verifier model call recorded 09:14:05Critical alert raised, person on tracks 09:15:10Alert acknowledged by an operator 09:16:00Clip downloaded by a reviewer 09:20:41Lesson proposed, not a hazard here 09:21:03Lesson accepted by an administrator 23:59:00Old evidence deleted, tombstones written

Runs on an edge AI gateway at your site.

We deliver the gateway, the rules for your use cases, and a proof of value on your own footage.

  • Edge AI gatewayOne desk-sized unit runs all three agents and the workspace. No cloud needed.
  • Field kitThe same software on a laptop with an NVIDIA GPU, for vehicles and temporary sites.
  • Any cameraDrones, robots, vehicle and body-worn cameras, recorded or live.
  • Your lookDay, night and wall modes, in your own brand if you have one.
no route out

See it on your own footage.

Send us a clip and tell us what to watch for. We run it on the gateway and show you every alert and finding.

hello@lenstat.com QA+974 6650 7955 AU+61 480 009 095 US+1 234 529 3634
What we need
  1. VideoA recording, or a live stream at your site.
  2. A flight log, if you have oneIt adds position and scale.
  3. The use caseWhich of the five, and the area to watch.