What the camera sees, written into the record by computer vision and OCR.

Every read carries a number.

Above your threshold, a field or a detection is written to the record. Below it, the read queues for a person, with the crop of the image beside the value being checked.

A read creates or updates a record and does nothing else. Paying, deleting and contacting a customer stay with a person.

Corrections feed back in. The fields people fix most often are tuned next, and a threshold goes up only after the correction queue shows the reads have earned it.

Six kinds of read, each ending in a record.

• Document capture. Invoices, purchase orders, delivery notes and statements read into structured fields and posted to the ERP.
• Identity and compliance reads. National ID cards, passports and trade licences parsed into the fields an onboarding form needs, and handled as the personal data they are.
• Arabic and mixed-script OCR. Printed text, and handwriting within limits, measured on your documents and reported in full, weak fields included.
• Counting and presence. Pallets, vehicles, stock on a shelf and people in a zone, counted continuously.
• Defect and condition checks. Damage, seal integrity, labelling and packing errors flagged while they can still be fixed.
• Tracking. One object followed from frame to frame, which is what makes counting, dwell time and direction measurable.

The limits, stated before you buy.

The table sets out the pattern to expect, so the weak spots are known before any money is spent. Your own accuracy is then measured per field and per document type, on your documents or footage, and you see that number before we commit to anything.

Footage stays on site, and only records travel.

A small machine on site runs the model, and the network carries the record alone. Your existing cameras may well do. Many IP cameras conform to ONVIF, the interoperability standard for IP video, and stream over RTSP, so the machine on site can read them as they are. ONVIF is retiring Profile S, its long-standing streaming profile, in favour of Profile T, and cameras already installed keep working. Placement and light come before the model. Where a camera cannot see the thing, we raise it before you spend money on software to compensate.

A detection says what is in this frame, and a track says it is the same object as in the last one. A track supports the one reliable prediction at this level: where an object will be in the next second or two, given where it has been. That is enough to raise an alarm before a collision, or to tell a vehicle leaving from one arriving.

We do not sell intent. A model can tell that a path continues. What somebody meant to do is beyond it.

ID documents are personal data.

An identity document is personal data before anything is done with it, and the law that governs it depends on where you operate. In the UAE, outside the DIFC and ADGM, which run their own data protection laws, the main federal law is Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data, and its definition of sensitive personal data takes in biometric data. Treat this as general information, and confirm it with your legal adviser.

Originals are kept as long as your policy says and deleted on schedule, each field is shown only to the roles that need it, and every read and every view is logged. Your legal adviser sets the lawful basis and the retention period, and we build to that decision.

A read captures what is printed on the card. Whether the holder is the person on it needs a separate check, which is the job of a national digital identity such as UAE PASS.

Built on your own bad scans and footage.

• Sample set. A few hundred of your real documents or a few hours of real footage. Clean samples prove nothing, so we ask for the bad scans.
• Baseline measurement. Accuracy per field and per document type, measured first. If the reads cannot clear the bar, you learn it at the start.
• Pipeline and thresholds. Extraction, validation against your own rules (totals add up, dates are possible, the TRN has its 15 digits), and a queue for everything below the line.
• Correction screen. The reviewer sees the image crop beside the value and accepts or fixes it from the keyboard.
• Integration. Into the ERP, the document store or the workflow, with an audit trail that can reconstruct any record.
• Tuning. Real corrections, week by week, once the system is live.

One fixed price, with the edge hardware named.

The build is quoted as one fixed total and paid in milestones. For camera work the edge hardware is named and priced in the proposal, and hosting and inference are estimated from your real volumes before you commit.

The source code, the model configuration and the evaluation sets pass to you on full payment. Your images never train a model outside your company, and the pricing page covers warranty and hosting.

For a first look, a still from the camera that matters and a few of your worst scans are enough. The call takes 30 minutes and costs nothing.