AI models that watch your sites — and turn what they see into incidents you can manage.
Your cameras, machines and network are already producing signals nobody has time to watch. NepalBix runs AI models on site that recognise what actually matters, raise it as an incident, put it in front of the right person, and keep the record afterwards.
Detection is the easy half. The hard half is what happens next.
Plenty of systems can tell you something moved. The reason most of them end up ignored is that nobody built the part where a finding becomes a job somebody owns, finishes, and can show an auditor a year later.
Models watch, continuously
Running on a device in your building, on the cameras and sensors you already have. Nothing has to leave the site for a model to reach a conclusion.
A finding becomes an incident
Not a line in a log. An incident with a type, a severity, the footage attached, and the site and camera it came from.
It reaches the right people
Routed by what it is and where it happened — to the console, to the channels your team already uses, and escalated if nobody picks it up.
Somebody works it and closes it
Acknowledged, reviewed against the footage, confirmed or dismissed, and resolved — with who did what recorded.
The record outlives the incident
Searchable months later, including everything the system decided was not worth raising.
Every incident carries its own severity, owner and state — from any of the three model families, in one list.
Three families of models, running side by side.
They look at completely different signals and produce the same thing: an incident, in the same inbox, graded the same way.
Models that watch
Trained to recognise people, vehicles, objects and body posture — and then to judge whether what they are doing matters.
- People where they shouldn't be
- Objects left behind
- Fire and smoke
- Vehicles among pedestrians
- Falls and altercations
- How busy a space is
Models that listen
They learn what normal looks like for each individual machine, at each way it is run — then notice when it drifts.
- Motors and pumps beginning to wear
- Bearings developing a fault
- Temperature, pressure and speed drifting
- A machine behaving unlike itself
- Readings taken from your controllers, read-only
Models that listen in
A passive view of the network your building runs on, looking for what has changed rather than what is famous.
- Devices that have never been seen here
- Traffic that doesn't belong on this network
- Entirely by listening — it never transmits
- Safe to run alongside live machinery
Eight things the vision models raise.
You choose which apply to each camera and draw the areas that matter — a doorway, a machine cell, a fence line — on the camera's own picture.
Someone crossed a line
A fence, a doorway, a threshold — including which direction they went.
Someone is lingering
In a restricted area, or anywhere they've stayed longer than they should.
A bag was left behind
Something set down and walked away from, still there minutes later.
Fire or smoke
Recognised from the picture itself — often before a ceiling detector would trip.
A vehicle where people walk
Forklifts, trucks and carts straying into pedestrian space.
Someone fell, or a fight
Judged from body posture, not just something moving oddly.
How many, and how busy
People in and out, how full a space is, where crowds build up.
Areas nobody may view
Blank out a window, a bed, a counter — permanently, for everyone.
Every finding becomes an incident with an owner and an ending.
This is the part that decides whether a detection system is still switched on in a year. A finding that nobody is accountable for is just noise with extra steps.
A person on a fence line at 3am and a queue getting long are both incidents, and they are not treated the same way.
If an incident isn't acknowledged it gets re-sent, and the serious ones escalate until somebody picks up — including a phone call.
The clip is attached to the incident, including the seconds before it. Confirming or dismissing takes one look, not a hunt through recordings.
Anything the models raised but the system decided not to escalate is still recorded, with the reason. Nothing disappears silently.
A system that cries wolf gets switched off.
Anyone who has run detection at scale knows this decides whether it survives past month two. So a finding has to earn its way to a person.
Each area has its own hours. A goods yard busy all day may only be interesting at night, and it won't page anyone at noon.
The same situation doesn't generate a stream of identical incidents — but a genuinely different person or event still does.
Before a borderline finding reaches anyone, a second model looks at the picture again to rule out glare, reflections and shadows.
Held-back findings are still recorded with their reason, so you can always audit what the system chose not to raise.
And then: what's actually going on across your sites?
Individual incidents tell you about a moment. The value over a year is in the pattern — which site, which shift, which door, which machine, getting better or worse.
Which sites are online, which cameras are healthy, which devices need attention — reported by the sites themselves, continuously.
Incidents by type, area and time of day. How busy a space gets, when, and whether that is changing week on week.
A camera that has quietly stopped producing a picture is itself an incident. You find out from the platform, not from a gap in an investigation.
Search the record by site, camera, type and date, and get the answer in seconds rather than requesting an export.
Models drift. This one is built expecting that.
A model that was accurate the week it was installed is not automatically accurate after the lighting changed, the line was rebuilt, or the camera was nudged.
When your team confirms or dismisses an incident, that judgement is captured — and it's what makes the next version better at your site specifically.
The platform tracks when what a camera is seeing has moved away from what its model was trained on, and says so.
A new model runs alongside the live one on your real footage and has to show what it would have done differently before anyone considers switching.
Model and software updates are approved by you, rolled out site by site, and reversed automatically if anything looks wrong.
The same platform, a different job in every building.
Pick the one closest to yours. The models don't change — which ones you switch on does.
Industrial plants & factories
Safety incidents recorded the moment they happen, and machinery watched for the failure that hasn't happened yet — without anything being connected to the controls that run the line.
Someone enters a machine cell
A high-severity incident is raised the moment it happens, with the footage, so supervisors respond in the minute rather than reading about it in a report.
A forklift crosses a walkway
Near-misses become a countable, trended thing instead of something people mention informally.
A motor starts to sound wrong
Wear surfaces as an incident for maintenance weeks before failure, so the repair is scheduled rather than unplanned downtime.
Fire in a production area
Recognised from the camera picture, often earlier than a ceiling sensor in a large open space.
An unknown device on the plant network
Something that has never been seen on that network is raised for IT — noticed purely by listening, with nothing sent onto the network.
A press behaving unlike itself
Readings drifting from how that specific machine normally runs at that specific setting. Readings only — the platform never writes back.
Airports
Large, busy, heavily regulated, and full of boundaries that matter. Most of the value is in noticing things in crowds quickly — and being able to prove afterwards what was noticed and when.
An unattended bag in the terminal
Raised with the footage of who left it and when, so the response starts from evidence rather than a description.
Entry to an airside or restricted door
Access boundaries watched continuously, including which direction someone passed through.
Queues building at security
Live counts per area, trended by hour and day, so staffing decisions are made on numbers.
Perimeter crossing after dark
Fence lines watched all night, with direction recorded so a vehicle leaving reads differently from someone arriving.
Baggage handling wearing out
Conveyor drives monitored so belts are serviced on a schedule rather than during a morning rush.
Something new on the operations network
Ground systems run on networks that shouldn't change quietly. Anything that appears is raised.
Hospitals & care facilities
Patient and staff safety, with privacy as a hard requirement. Sensitive areas are blanked out permanently, and every incident carries the record an investigation will later ask for.
A patient falls in a corridor
Raised immediately, day or night, including in areas nobody is watching a monitor for.
Unauthorised entry to a drug store
Restricted rooms watched continuously, with the clip attached for the incident record.
Aggression in the emergency department
An altercation starting is raised to security while it's still an argument.
Waiting rooms filling up
Occupancy per area, trended, so crowding is managed before it becomes a complaint or an infection-control problem.
Backup generators and chillers
The plant nobody thinks about until it fails, watched continuously so it doesn't fail unannounced.
Connected medical equipment
Devices appearing on clinical networks are raised for review — passively, never by probing the equipment itself.
Hotels & hospitality
Guest safety and back-of-house security, without the system becoming something guests would object to. Guest-facing areas can be excluded entirely.
Loitering at a staff-only door
Back-of-house entrances watched, so an unfamiliar person waiting by a service door is raised to duty staff.
Luggage left alone in the lobby
Raised early and quietly — usually a forgetful guest, occasionally not.
Smoke in a service area or kitchen
Early warning from cameras in spaces where a small fire can go unnoticed for minutes.
Check-in queue length
How many guests are waiting right now, and at which hours it repeats, so a second desk opens on evidence.
Lifts, pumps and air handling
Building plant monitored so a failure becomes a scheduled repair rather than a guest complaint.
Rooms and windows permanently blanked
Any area you decide is off-limits is blacked out before a person or a recording can ever see it.
Embassies & government buildings
Perimeter-first, evidence-grade, and built for premises where footage must not leave the site unnecessarily and every decision must be accountable.
Someone crosses the perimeter
Boundary lines watched continuously, with the direction of crossing recorded on the incident.
A person waiting outside the gate
Loitering over time along a boundary, rather than a single person walking past.
An object left at the entrance
Anything placed near an entry point and abandoned is raised immediately, with footage of the approach.
Movement in a restricted wing
Internal boundaries treated as seriously as external ones, with an auditable record of every incident and every response.
A device that doesn't belong
The building network watched for anything new appearing on it, purely by listening.
Everything kept as sealed evidence
Clips are encrypted, released only on request, and the full decision trail behind every incident is preserved.
Warehouses & logistics
Fast-moving vehicles and people sharing a floor, doors that should be shut, and equipment that only gets attention when it stops.
Forklifts and people in the same aisle
The most common serious injury in a warehouse, counted and trended rather than audited after the fact.
A pallet blocking a fire exit
Raised while it's still easy to move, and recorded for the safety file.
A dock door left open after hours
Boundaries that should be shut at night, checked without anyone walking the floor.
Conveyors and sorters
Drive wear surfaced early, so maintenance lands between shifts instead of during peak.
How busy each zone is
Where people and goods actually accumulate, over a shift, a week, a season.
Unknown devices on the site network
Scanners, printers and handhelds change often — anything genuinely new is raised.
Privacy is handled at the camera, not by policy.
For hospitals, hotels and diplomatic premises this usually decides whether a system is permitted at all. So it isn't a setting somebody can quietly switch off.
Areas you mark private are obscured on the device itself — before the picture reaches a screen, a recording or an incident. Not hidden afterwards: never available.
A neighbour's balcony, one bed in a ward, a reception counter — masked as the exact shape you draw, so you don't blank half a room to cover a corner.
Video is processed on your site. Clips are sealed, and released only when somebody asks for a specific one.
Every incident, and every finding deliberately held back, is logged with its reason — which is exactly what an auditor or an inquiry asks for.
One building or two hundred, one screen.
A chain of hotels, a group of hospitals, a manufacturer with plants in four countries — same console, one list of incidents.
The device is plugged in, identifies itself and pulls its own settings. No specialist visit to configure it.
It restarts its own parts if they fail, reconnects dropped cameras, and reports its own health so you find out before your staff do.
If the connection drops, the models keep running and incidents keep being recorded locally. The console catches up when the link returns.
Sites on weak or metered links send far less data than streaming video would, and urgent incidents are never held back to save bandwidth.
Getting started is a day, not a project.
A pilot is one device, your existing cameras, and one or two areas that matter.
We look at what you already have
Existing cameras usually work as they are. Coverage gaps, if any, are identified up front.
The device goes in
One unit, in an existing cabinet, on the network your cameras are already on. Nothing is connected to operational controls.
We mark up the areas together
Where the boundaries are, what counts as an incident, which areas must be blanked out entirely.
You decide who gets what
Which incidents go to whom, at what severity, on which channels, and how they escalate if nobody answers.
It runs, and you tune it
The first weeks are about making sure every incident is worth reading. After that it mostly gets out of the way.
See it on your own cameras.
One device, one area marked up, and a week of your real footage is enough to know whether it earns its place.