Your cameras already see everything. CamBrain is the part that understands it.
Everything local. Everything private. The AI runs on one box inside your building — nothing is uploaded, nothing is shared.
Not a wall of alerts. One ticket, one owner, one clock — and a re-inspection that decides whether it was really fixed.
Every box below was produced by running our detectors over dashcam footage from Exhibition Road. Nothing is staged, nothing is drawn on afterwards.
Garbage · encroachment · hoardings · potholes · vehicle and pedestrian tracking — one pass.
We read your existing streams over RTSP or ONVIF. Nothing is replaced, rewired or re-pointed. Read-only — CamBrain cannot move your cameras or change your recordings.
One box on site runs every detector in a single pass over the frame: crowd, faces, behaviour, civic faults, camera health. The video never leaves the building.
Detections are not incidents. Aggregation, deduplication, hysteresis and cooldown turn hundreds of sightings into the few that deserve a person's attention.
Every ticket gets an owner and a clock. When someone marks it fixed, CamBrain re-inspects the same place on fresh footage and reopens it if the problem is still there.
Density at a gate, queue length at a counter, occupancy against a limit — and people pushing against the flow, which is the first sign of a crush minutes before density peaks.
Detection and counting with no identity attached. Watchlist matching is available, and ships only under conditions we will not waive.
Logic over tracks and time, not a black box guessing at intent. Every verdict carries the measurement behind it.
Surveyed from a dashcam on a route, or watched continuously from a pole. Sized against the threshold in your contract.
The unglamorous one that makes every other number trustworthy. Across a large estate a fraction of cameras is always blind, and nobody reports it.
The layer no open-source project ships, and the reason a finding becomes a fix rather than a notification.
CamBrain is a layer, not a replacement. If your cameras produce a stream, we can read it — whatever badge is on the housing.
Nothing streams to a cloud, so the compute has to be on site. These are the three configurations we deploy, priced as a monthly rental with the hardware included.
Video storage stays on your existing NVR. The box keeps events, thumbnails and short clips — not your archive.
Under the DPDP Act the awkward question is not whether the technology works. It is where the footage went, who looked at it, and why. On-premise makes most of that answerable in one sentence.
Every capability carries four numbers — accuracy, false positives, latency and the hardware it ran on. Where a detector does not transfer to local footage, we say so and show the measurement rather than quietly dropping it from the list.
Our road-damage detector scores 0.306 on the international benchmark and 0.001 on our own Patna street footage. It does not transfer to that camera angle yet. We measured that ourselves, on 300 frames we annotated by hand, and it is on this page because you would rather hear it from us.
No. If a camera produces an RTSP or ONVIF stream, CamBrain can read it. Older and lower-resolution cameras reduce accuracy at distance, and the site survey tells you which ones before you commit.
No. Analysis runs on a box inside your building. The only thing that crosses your network boundary is the dashboard connection, and the system keeps full function with the internet unplugged.
Not yet, and we will not claim otherwise. At our shipped operating point roughly one in five encroachment reports is wrong. That is fine for a queue an officer reviews and it is not fine for an automatic fine. Every number is published so you can judge for yourself.
Every finding carries the evidence frame and the measurement behind it, so a reviewer sees what the machine actually saw rather than a verdict. Where the system cannot tell, it returns UNCERTAIN instead of guessing — "we could not measure this" is a first-class answer.
We can detect and count faces with no identity attached, and that is the default. Watchlist matching requires accuracy measured per skin tone, human confirmation before any action, per-officer query logging and a documented legal basis from the authority. We do not enable it without all four.
A monthly rental per site with the hardware included, sized by camera count. No capital purchase, and the box stays ours — which matters when a contract ends.
Every dependency is recorded with its licence and the provenance of its weights — including the ones we rejected and why. AGPL detectors and non-commercial model weights are excluded from the product. We can answer a procurement question about this in one sentence.
Read-only access to a single existing stream is enough to show you what your own estate is already recording — and what nobody is watching.
CamBrain AI · cambrain.ai · runs on-premise, offline · Patna, Bihar