Will this work with our existing CCTV cameras?
In almost all cases, yes. We connect to any camera exposing an RTSP, ONVIF, HTTP or SDK stream, which covers the overwhelming majority of IP cameras, NVRs and DVRs in service today. Analogue cameras running through a networked DVR also work. You do not need to replace hardware to start — and if a specific detection genuinely needs a better angle, we identify it during the free audit rather than after you have committed.
Do we need a GPU, or will a normal server run it?
It depends on the algorithm class. Basic analytics — people counting, line crossing, intrusion, loitering, dwell time, occupancy — run acceptably on a Core i7 class CPU for a modest number of streams. Advanced analytics — multi-pattern shoplifting behaviour, pose-based fall and fight detection, PPE compliance, face recognition, cross-camera re-identification — require an NVIDIA RTX class GPU with 8 GB VRAM or more. We size hardware during the feasibility audit against your actual camera count and feature set.
Should we deploy on-premise or in the cloud?
On-premise suits sites with many cameras, limited upload bandwidth or strict data residency rules, because video never leaves the building and detection continues through an internet outage. Cloud suits multi-site and franchise operators who want one dashboard across locations with no server to maintain. Hybrid is the most common outcome in practice: inference runs on a local edge server, and only events, metadata and short evidence clips sync to the cloud.
Why does camera calibration matter so much?
Because accuracy is set by geometry, not only by the model. Every camera is calibrated to its premises — mounting height and angle, lens distortion, real-world scale so distances and speeds mean something, lighting and glare across the day, and the zones that carry business meaning such as tills, entrances, shelves, workstations and restricted areas. Calibration is specific to each site, each camera and each use case; it does not transfer between them. Uncalibrated deployment is the most common root cause of false alarms in this field.
How accurate is AI shoplifting detection?
Our behaviour engine recognises more than 500 distinct shoplifting and concealment patterns and returns a confidence score per tracked person rather than a binary verdict. Real-world accuracy depends heavily on camera placement and calibration, which is why we always run a pilot on your own site and tune thresholds against your own footage and your own tolerance for false positives before any rollout. Any vendor quoting a single accuracy percentage without seeing your cameras is quoting a benchmark, not a result.
Can it integrate with our POS, ERP or access control?
Yes, and for several use cases it is essential rather than optional. Detecting goods leaving unpaid, register fraud, order handover errors or estimated financial impact requires correlating video events against transaction data — a camera alone cannot distinguish "walked out without paying" from "paid at the other till". We integrate with POS, ERP, WMS, HRMS, access control and alarm panels via API, database connector or scheduled file exchange.
How long does a deployment take?
A single-site pilot covering a defined detection set typically runs four to eight weeks from audit to live alerts. A full multi-site platform with dashboards, mobile apps, role-based access, integrations and billing typically runs three to six months to first production release, then rolls out site by site using a repeatable calibration playbook.
Is face recognition required, and is it legal where we operate?
It is optional, and most deployments do not need it — anonymous person tracking is sufficient for theft, safety, queue and productivity analytics. Where identity is genuinely required, such as staff attendance, we build it as a separately controlled module with consent capture, retention limits and audit logging, and we scope it against the biometric and privacy law in your jurisdiction before design begins. Several regions restrict or prohibit biometric processing and workplace inference; we would rather raise that in week one than in the compliance review.
How do you keep false alarms under control?
Four layers. Per-camera calibration and business zones so events only fire where they matter. Confidence thresholds tuned on your own footage rather than a public benchmark. Temporal confirmation, so an event must persist across frames before it escalates. And an operator review loop where false positives are marked in the dashboard and fed into the next tuning cycle. Alert fatigue kills these systems faster than model error does, so we optimise for precision first and add breadth afterwards.
Can we white-label the platform and resell it?
Yes. We build multi-tenant, white-label-ready platforms with your branding, your domain, role-based permissions, tenant provisioning and subscription billing, so system integrators, security firms and SaaS operators can take it to market under their own name with Starter, Professional and Enterprise tiers.