How CamBrain AI handles video, access and evidence.
What follows describes how the product actually works. No system is perfectly secure, and no AI is always right; these are the controls we have and the lines we hold.
Topic by topic.
Where video is processed
On a server at your premises. Video is analysed there, not sent to a cloud AI service, and raw video is not uploaded to us. The server keeps an outbound, encrypted connection to our support hub so we can monitor and update it, and so your staff can view cameras remotely.
How streams are accessed
Read-only, over RTSP or ONVIF, from cameras directly or through your NVR or VMS. CamBrain AI cannot move cameras, change settings or alter your recordings.
Who can view footage
Only signed-in users with a role you grant. Remote viewing by your own staff travels over an encrypted connection through our relay server and is not stored on the way. Every view of recorded footage is logged against a named user.
How alerts are generated
From the use cases configured for each camera: a zone, a schedule, a count limit or a camera-health check. Detections are grouped, de-duplicated and given a cooldown, so one event produces one alert. An alert is a prompt for a person to look, not a finding of wrongdoing.
How evidence is handled
Snapshots and clips are fingerprinted with SHA-256 when they are captured or exported, so any later change can be detected. Exports are logged.
What gets logged
Sign-ins, views of recorded footage, exports, searches (with the user and stated reason), changes to rules and zones, and every acknowledgement and closure of an alert.
What is retained, and for how long
Defaults, adjustable per site within your policy: search index, unattached frames and, where face memory is enabled, face templates 14 days; crowd readings 21 days; plate reads 30 days; door and visit events 90 days; audit logs kept for at least 180 days. Evidence attached to an alert is kept until an administrator deletes it — automatic expiry 365 days after closure is planned, not yet enforced. Recordings roll over as storage fills, on a period you set.
What needs specific authorisation
Search for a person, missing-person cases, number-plate watchlists and any face matching. These run only for authorised users, with every query logged against a named person and a reason, and every match confirmed by a person before anyone acts.
What we will not build.
Put in writing before a project starts, because an administration or a company is entitled to know where a vendor's line is.
- Untargeted tracking of individuals across a camera network
- Emotion detection, gait recognition or "suspicious behaviour" prediction
- Identifying organisers or participants of any lawful assembly
- Worker productivity scoring, or shopper age, gender or demographic profiling
- Cameras in bedrooms, bathrooms, wards, patient rooms or examinations
Only with conditions.
Face matching is not offered for sale until four conditions are met. One is not met yet: accuracy by skin tone and lighting has not been measured. At Gaya, photograph searches for missing persons also compared faces seen by the cameras (face templates kept 14 days), under the administration's authority, with every match confirmed by an officer. The conditions:
A legal basis
Documented, and supplied by the authority.
Published accuracy
Measured per skin tone and lighting condition, not one headline figure.
A person decides
A named officer confirms every match before any action.
Every query logged
Against that officer, with a stated reason.
We trace the weights, not the badge.
Many public video datasets forbid commercial use, and a model trained on one carries that restriction into your deployment. Every model and dependency we ship is recorded with its licence and provenance — including the ones we rejected, and why. Ask us for the register.
Questions from your IT or legal team?
We will walk them through the architecture, data flows and retention for your site.