The challenge
Most CCTV is recorded but never really watched. The systems that do send alerts cry wolf constantly — wind, headlights, a passing cat — so operators stop trusting them, and real intrusions slip through. The usual fix is more guards, which is expensive and still blinks, sleeps, and takes breaks. And modern surveillance raises a real privacy question: who is watching the video, and where does it go?
Prahari set out to solve all three at once: cut false alarms to almost nothing, do it at a fraction of the cost of more guards, and keep raw footage private by design.
What we built
GMK Group built Prahari end to end — the computer-vision pipeline, the edge software, the rules engine, and the operator experience. It works with the cameras a business already has: point an edge box at the existing RTSP streams and it's live, no rip-and-replace. The heart of it is a five-stage cascade that filters noise so only verified intrusions ever reach a human:
- Motion gate — roughly 99% of frames stop here; only meaningful movement moves on.
- Fast detector — a YOLO model identifies person, vehicle, animal, or bag in the frame.
- Rules engine — intrusion, tripwire, loitering, vehicle-stopped — tuned to each site's rules.
- Nuisance suppressor — per-camera learning, so false alarms fall week over week.
- AI verification — a plain-English description of the scene plus a recommended action reaches the operator.
How we engineered it
Two constraints shaped the architecture from day one: privacy and cost. Prahari is edge-first — continuous video never leaves the site; only short event clips, a few per night, are sent to the cloud. There is no facial recognition, ever, and the system is built to be DPDP-compliant. That means a business gets the alerting it needs without turning its premises into a surveillance dragnet.
On cost, the goal was to make automated monitoring dramatically cheaper than adding headcount — on the order of a tenth of the monthly cost of a single extra guard post, while watching far more than any one person could. The AI verification layer is what makes that trustworthy: instead of a raw "motion detected" ping, the operator gets a clear sentence — "one person has climbed the perimeter fence and is moving toward the yard" — and a recommended action. This is the bar GMK builds to: production systems that hold up when something real happens, because a security alert can't wait for a bug fix.
The outcome
Prahari is live and available, with a free trial and simple per-camera pricing. The result for an operator is fewer false alarms, real intrusions that still get through, raw video that stays on-site, and monitoring at a fraction of the cost of more guards — all on top of the cameras they already own.
Visit the live product → praharisecurities.com
Have an AI product you need built — and built right?
We design, build, and ship production AI and software end to end. Tell us what you're trying to build.
Start a conversation