Frequently Asked Questions
The questions we hear most often on video analytics projects, with straight answers. If your question is not here, write to us.
General
What does CX Teknoloji do?
CX Teknoloji is a computer vision company that uses AI to turn the footage of existing IP/CCTV cameras into events and measurable data. It was founded in 2021 and is headquartered at GTU Technopark in Gebze, Türkiye. It offers 23 ready solutions across 5 categories: occupational health and safety, zone and perimeter security, production efficiency (OEE/MES), retail analytics and industrial machine vision. Needs the ready solutions do not cover are met with custom video analytics. The full list is on the Solutions page; company and team details are on the About page.
Which industries do you serve?
Manufacturing and factories, warehousing and logistics, energy and hazardous-materials facilities, construction sites, retail stores, and quality control on production lines. Whether a site is in scope depends not on its industry but on whether a measurable event or metric can be extracted from its camera footage. Industry-specific scenarios are collected on separate pages (in Turkish): shipyards, chemicals, organized industrial zones, automotive, logistics, ports, metal casting and food and pharma.
What is EscherSense, and how does it relate to CX Teknoloji?
EscherSense is CX Teknoloji's retail brand. It covers the store analytics solutions: people counting, demographic analysis, heatmaps and queue management. It uses the same inference pipeline as the industrial solutions; there is no separate software or team. Its canonical address is cxteknoloji.com/eschersense/.
Where are you located, and which regions do you serve?
Our office is at GTU Technopark, Gebze / Kocaeli. We have separate pages (in Turkish) for the industrial regions of the Marmara area: Gebze, Kocaeli, Istanbul, Bursa, Sakarya, Tekirdağ and Yalova. Beyond these, we run projects across Türkiye and in Europe; we work in Turkish and English. The site survey is done on site, and the analytics server is installed at your facility.
Technical
Does it work with our existing cameras?
At most facilities, yes: no additional camera investment is needed, and a GPU server is installed on site for processing. IP cameras that provide an RTSP stream or are ONVIF-compliant connect to this server directly or through the NVR/VMS. Suitability depends less on the brand than on angle, resolution, bitrate and night lighting. During the survey each camera is put into one of three groups: usable as is, usable after an angle or lens change, and needing a new position. In-line quality inspection and some licence plate recognition points usually need industrial or specially positioned cameras. The camera suitability assessment is free. More: ONVIF and VMS integration (in Turkish).
How many cameras can one server handle?
A single GPU server typically processes 8–16 cameras in real time, depending on the solution type. With heavy models such as pose analysis this drops to 6–10, because the pose model computes 17 keypoints per person. The exact number depends on stream resolution, frames processed per second, model weight and the number of scenarios running on the same camera; scenarios that can work on the sub-stream put less load on the server. Capacity is calculated during the survey using the real streams on site, with headroom left for cameras and scenarios added later. More: system architecture from RTSP to edge GPU (in Turkish).
What is your accuracy rate?
There is no single rate; accuracy varies by solution, camera placement and site. In our reference projects, 96% vehicle and equipment safety and 96.2% counting accuracy were measured. The value for your site is measured in the pilot: what counts as a correct detection is written down up front, and real events are labelled and compared with the system's detections. Precision (how many alerts were real) and recall (how many real events were caught) are reported separately. Because accuracy is tracked per camera, a low-performing point can be traced to either placement or the model.
How do you reduce false alarms?
With a three-layer filter. A confidence threshold drops uncertain detections; a rule engine intersects each detection with a defined zone, direction and time; temporal validation waits for the condition to persist over consecutive frames, after which repeat suppression is applied. Thresholds are calibrated per camera and object class on your site during the pilot; site-specific false alarm examples are labelled and the model is retrained on them. Because objects are recognised instead of motion, and people, vehicles and animals are separated, the false alarm reduction of our perimeter security solution is published as −85%. More: reducing false alarms (in Turkish).
Edge (on-premise) or cloud?
The default installation is on premise: footage is processed on the GPU server at the facility and raw video is not sent to the cloud. Processing can also run in the cloud if preferred. On-premise processing has three benefits: response time stays predictable because decisions are made on the local network, only event records and short clips use the internet connection, and analysis and local alarms keep running if the internet goes down. Multi-site organisations often use a hybrid setup, with inference on site and dashboards and model updates at the centre. If company policy requires it, no cloud layer is set up at all. More: edge vs cloud (in Turkish).
Does the system keep working if the internet connection drops?
With an on-premise installation, yes. Detection, the rule engine and local alarms run on the server at the facility, so alerts such as sirens, IoT relays or the fire alarm panel do not depend on the internet. During an outage, events bound for the centre are buffered locally and synchronised when the connection returns; only remote dashboard access and multi-site reports are affected. An uninterruptible power supply is recommended for the edge server. Deliberately cutting the internet during the pilot and checking that the buffered events all arrive is a good test.
Which systems can it integrate with?
Events are published over REST API, webhook and MQTT, and connect this way to MES, ERP/SAP, WMS, SCADA/PLC, VMS, POS, CRM and occupational health and safety software. Every event carries a timestamp, the camera or line/station ID, the event type and a clip reference; no raw video is sent to the receiving system. On-site alerts are delivered via IoT alarm relay, siren, fire alarm panel, email/SMS, WhatsApp and mobile notification. On sites that use a VMS, the event appears as an alarm on the operator screen and the relevant camera opens. More: MES integration (in Turkish).
Will our existing recording system (NVR/VMS) be affected?
Not when installed correctly; the recording system keeps doing its job. The analytics runs on a separate server and usually uses the camera's second stream, leaving the main stream for recording. Because cameras can serve only a limited number of simultaneous sessions, this is checked for each model during the survey; if needed, the stream is taken from the VMS instead of the camera. Cabling stays in place and the cameras' software settings are not touched; if an angle change is needed, it is proposed with its reasoning during the survey. If the connection is removed after the pilot, the site returns to its pre-pilot state.
What are the minimum camera requirements?
For most scenarios a 1080p IP/CCTV camera is enough; the real criterion is the number of pixels on the target. For person detection the body should appear at least 60–80 pixels tall, for PPE tracking at least 80 pixels, and for pose and safe-behaviour analysis at least 120 pixels. This value results from distance, lens focal length and sensor resolution together; a wide-angle 4K camera can give fewer pixels on a distant target. The angle also depends on the scenario: top-down for people counting, a 20–30 degree tilt for PPE, parallel to the road for plate reading. Night is assessed separately because IR footage loses colour information. More: camera placement guide (in Turkish).
Process and pricing
How long does installation take?
For ready solutions, installation is typically completed within a week, followed by a 30-day pilot: the first week is for calibration, and the rest measures real precision and recall. Custom models add feasibility, data collection and labelling, training, site testing and commissioning; a typical custom project goes live within 12–14 weeks. Cameras and the recording system keep working throughout the installation; if an angle change is needed, it is done at a time planned with the site team. The project-specific timeline is shared in writing after the survey. More: installation process (in Turkish).
How does the pilot work?
The pilot runs for 30 days on one problem, one area and at most 8 cameras. Two limits are agreed in writing before it starts: acceptable accuracy, and the number of false alarms an operator can realistically review per shift. The definition of a correct detection and the out-of-scope cases are in the same document. On site, polygons, thresholds and time windows are tuned, and false positives are reviewed daily. At the end of the pilot, one of three decisions is made based on the measured values: roll out, revise and run another round, or stop. More: designing the right pilot (in Turkish).
How is pricing determined?
Pricing is per project; a single per-camera unit price does not distinguish between different projects with the same number of cameras. The items that shape a quote are: the number of cameras and the usability of existing ones, the number of scenarios and model weight, server hardware, where processing runs (on premise by default, cloud if preferred), the licence model, installation and calibration, integration scope, custom model training if needed, and maintenance. We recommend requesting quotes against the same item list and comparing licence and subscription over a three- or five-year total. The camera suitability assessment and the 15-minute demo call are free. More: AI camera system cost.
If there is no ready solution, do you develop a custom one?
Yes. For needs the ready solutions do not cover, such as a defect type seen only on that line or a movement that matters only at that facility, we develop custom video analytics. The process consists of feasibility, data collection and labelling, model training, site testing and commissioning; a typical custom project goes live in 12–14 weeks. Feasibility is done with 1–2 weeks of sample footage from the site; if the target cannot be measured with the existing footage, a camera or lighting change is proposed. The custom model runs on the same server and the same event management layer as the ready solutions.
How can I get a demo?
Reach us through the Contact page, at +90 549 774 86 46 or at info@cxteknoloji.com. In a 15-minute call we identify which solution can be tested with your existing cameras. None of it is required, but the following speeds up the survey: the problem you want to solve, a camera list (brand, model, location), recorder or VMS details, who should receive notifications and, if possible, a few hours of sample footage from the relevant cameras covering both day and night. The next step is the free camera suitability assessment.
Data and compliance
Where is the data processed, and is it KVKK-compliant?
The installation is configured to comply with KVKK (Türkiye's data protection law) and the GDPR; the final assessment is made by your organisation, as data controller, and your legal team. On the technical side, footage is processed on the on-premise server by default and raw video is not sent to the cloud (processing can also run in the cloud if preferred). The system records events, not identities: no facial recognition, no biometric records, and event clips are deleted automatically after a defined period. On your side, the privacy notice, the VERBİS registration and access rights are updated for the new purpose. According to the Turkish Personal Data Protection Authority's announcement of 8 June 2026, monitoring employee productivity is not a legitimate purpose. More: KVKK-compliant video analytics (in Turkish) and workplace camera recording and KVKK.
Do you do facial recognition?
No. Our solutions detect people, objects, poses and behaviour; no identification is performed and no biometric records are kept. A person is tracked across frames with a temporary, anonymous track ID. In retail, demographic analysis classifies age band and gender statistically and reports only hourly and daily distributions; no face images are stored. Because it needs a near-face angle, it is advisable to describe it explicitly in the privacy notice. KVKK treats biometric data as a special category and subjects its processing to strict conditions; event detection does not need such data.
How long are event records kept?
The period is set by your organisation's retention and destruction policy, and the system applies it: expired clips are deleted automatically and the deletion is logged. KVKK does not set a single period; figures such as 15 or 30 days are not a general legal rule, and records are kept for the shortest time the purpose requires. The analytics does not keep raw recordings, which stay in your existing recording system; what is stored are event clips (e.g. 10 s before the event for a zone violation, 10 s before and 30 s after a fall). By default clips stay on the on-premise server, or in the cloud of your choice if preferred; faces in clips can be blurred on request.