Custom Video Analytics
Models built from scratch for needs that off-the-shelf solutions cannot meet.
Every site has a problem of its own: a defect type that exists only on that line, a movement that means something only in that facility.
For such needs we start with feasibility: we state plainly what can be measured with the existing footage, what accuracy is realistic and what the labelling effort costs. Data collection, labelling, model training, field testing and go-live follow — a typical custom project goes live within 12–14 weeks.
How it works
Is it measurable?
Samples of your existing footage show whether the target is technically achievable.
Collection and labelling
Representative data is collected on site; rare cases are sought out specifically.
Test and calibration
Accuracy is measured during the pilot and thresholds tuned to the site.
Accuracy is a target; how it is reached is a commitment
In custom projects the crucial point is defining up front how accuracy is measured. What counts as a “correct detection”, how many false alarms are acceptable and which dataset the test runs on are written into the contract. The curve below shows the path a typical custom model follows during a pilot.
| Feasibility and scoping | 2 weeks |
| Data collection and labelling | 4 weeks |
| Model training | 3 weeks |
| Field test and calibration | 3 weeks |
| Go-live and user training | 1 week |
Data, not the model, takes the largest share of the timeline; capturing rare cases determines project quality.
| W1 | 71 |
| W2 | 76 |
| W3 | 82 |
| W4 | 86 |
| W5 | 89 |
| W6 | 92 |
| W7 | 93 |
| W8 | 94 |
The curve rises as each week’s false detections from the site are added to the training data.
Use cases
Surface defect
Detection of line-specific defect types.
Produce grading
Separation by size and colour.
Passenger behaviour
Station-specific flow analysis.
Equipment state
Reading visual gauges and flagging signs of faults.
Technical requirements
| Start | Feasibility on 1–2 weeks of sample footage from your site |
| Data | Labelled by our team, with your approval |
| Hardware | Custom camera/lighting design included where needed |
| Delivery | Model, panel, API and user training |
| Maintenance | Periodic retraining and accuracy reporting |
Frequently asked questions
How does Custom Video Analytics work?
Every site has a problem of its own: a defect type that exists only on that line, a movement that means something only in that facility.
What results and indicators does it provide?
Key indicators: Typical time to go live: 12–14 weeks; Feasibility → training → test → live: 4 stages; Site-specific training data: 100%. Data, not the model, takes the largest share of the timeline; capturing rare cases determines project quality. Thresholds are calibrated on your own site during the pilot.
Is it compliant with KVKK, Türkiye's data-protection law?
Footage is processed on an edge server inside your facility; raw video is not sent to the cloud. Facial recognition is not used; analytics outputs are anonymous and event clips are deleted automatically when the defined retention period ends.