Skip to content
Telesurety
Platform · 02 · Document intelligencePartly live

The platform reads the document and asks you when it isn't sure.

Every field comes with a confidence score and its location in the document. A field below the threshold is never written to the record on its own; it goes to human review.

Request a demo

Highlights

Registration reading
Live
Runs in the panel through the 'Fill from registration' button, in quote creation and when adding a vehicle on the customer card.
High-confidence precision
85.7%
Measured on the critical fields of real registrations, in a small set of documents.
Registration images
In-house
Images are read by an in-house service, never sent to an outside service and not stored.

The platform reads vehicle registrations, price comparison documents, and fuel and EV-charging invoices. A value read from a registration reaches the form, and an offer read from a price comparison reaches the customer, only after a person approves it. Registration reading runs entirely in-house.

How it works

Vehicle registration reading is live, with 85.7% field accuracy measured on real registrations. Insurers' price comparison documents are read in the panel; claim files are on the roadmap.

  1. 01 / 05

    Upload the registration

    In the panel, press 'Fill from registration' in a vehicle branch during quote creation, or when adding a vehicle on the customer card. Upload a photo or PDF of the registration.

  2. 02 / 05

    QR code first, then text

    On newer registrations, the document's own QR code is decoded first and supplies the document serial, the plate and the owner's identity or tax number. The code is used only when its content matches the registration's exact format, so another code in the frame cannot interfere. The model fine-tuned on Turkish registrations then reads the printed fields.

  3. 03 / 05

    Every field passes two checks

    Each value is checked against a fixed format rule. A field that obeys its rule and clears the model confidence threshold arrives as high confidence; if either is missing, it carries a 'Check' badge. If only a few critical fields can be read, the whole page is rejected and no field is suggested.

  4. 04 / 05

    Confirm by looking

    Read values appear in an editable table, side by side with the image snippet cut from that spot on the registration. The form receives plate, chassis number, document serial, make, trade name and model year. The owner's name and identity or tax number appear only in an information row for checking and are never written to the form.

  5. 05 / 05

    Apply and match the vehicle

    Press 'Apply to form' and the values move into the form. In quote creation, the plate and chassis are compared with the customer's registered vehicles: a match is selected, and with no match a pre-filled add-vehicle form opens. If the plate matches but the chassis does not, no vehicle is selected and a warning appears.

  6. Illustrative. Names and values shown are examples. Step by step. Upload the registration, QR code first, then text, Every field passes two checks, Confirm by looking, Apply and match the vehicle.

Measuring confidence

Live

High confidence means two independent signals agree.

The reading service maps the printed fields of the registration document, including engine number, colour, fuel type, seat count, weights and usage purpose. The panel form uses only the ones a vehicle record needs. A format rule is a hard signal: a value that breaks it is wrong however sure the model is. Model confidence is a softer signal, and its threshold is set from measurement. The label is a measurement; a person gives the approval.

  • A chassis number must have the correct length and contain no I, O or Q.
  • Plates must match the Turkish plate format, and the document serial must follow its own letter and digit pattern.
  • Identity and tax numbers must pass their checksums, and dates must be real calendar dates.
  • In tests on real registrations, the QR code corrected serial letters that the text reader had returned wrongly with high confidence.
  • The model is a text-recognition model fine-tuned on Turkish vehicle registrations. It replaced an earlier general-purpose attempt.

Quote and invoice documents

Live

A price comparison is tied to its request before it is published.

The panel reads the price comparison PDF prepared for a customer's quote request. It extracts the insured, the identity number after a checksum test, plate, product, coverage dates, preparation time, validity in days, and each offer row with its insurer and gross premium. Amounts are understood in both US and Turkish number formats.

  • If the identity number, plate or product does not match the request, the offers are not published to the customer app.
  • If no offer row or no validity period can be read, publishing is blocked. Insurer names are matched automatically, and staff can change the match.
  • A row whose amount cannot be interpreted with certainty is left out with a warning. The aim is that no price at the wrong scale reaches the customer.
  • At publish time the PDF is read again; the values seen in the preview are not relied on.
  • The company-vehicle module reads fuel and EV-charging e-archive invoices: plate, quantity and unit, unit price, line amount, VAT, invoice date or charging start, invoice number and seller. The plate is matched to the company vehicle, and an invoice number already entered is rejected as a duplicate.

Roadmap

Roadmap

Layout-independent reading and claim documents.

The method built for registrations is planned to extend to policies, quotes, driving licences, IDs, invoices and claim documents. Layout analysis will work alongside a small in-house language or vision-language model fine-tuned with LoRA, with output held to a fixed schema. Today's rule-based readers and a prototype that answers questions over policy PDFs with a local model will serve as baselines for measuring new models.

  • Claim files today have an expected-document list edited in settings, with document upload and preview. Automatic reading of accident reports, expert reports and invoices, and missing-document detection, are not built yet.
  • Confidence calibration is a research topic of its own: temperature scaling, conformal prediction and per-field thresholds. On real registrations, raising the threshold does not yet buy higher precision.
  • Document intelligence will become a stateless service that writes to neither disk nor database. For each field it will return the value, a confidence score, its location on the document and the model version.
  • Documents containing personal data will be processed only by in-house models on servers in Turkey. Cloud models will be used only with anonymous or synthetic data.
  • Every user correction will be logged. A model version that lowers the touchless record rate on the evaluation sets will not be released.

Use cases

From an intermediary's day.

  • 01

    A motor own-damage request with a registration photo

    A customer sends a photo of their registration for a motor own-damage quote. The agent presses 'Fill from registration' on the quote screen and uploads it. The chassis number and document serial arrive with the snippets cut from the registration, and the agent checks and approves them. Plate and chassis match a registered vehicle, so it is selected for the quote, and the agent sends the quote request to insurers.

  • 02

    Sending a price comparison to the customer

    A price comparison PDF for a motor liability request is uploaded to the panel. The panel extracts the offer rows with insurer and gross premium and confirms the document belongs to this request. The agent corrects one automatically matched insurer name and publishes the offers to the customer's app.

  • 03

    Month-end fuel records for company vehicles

    Fuel and charging invoices for the agency's company vehicles sit in one folder with subfolders. The person in charge hands the folder to bulk import. PDF invoices that are read and matched to a vehicle are added; unreadable or unmatched files stay in the list for fixing. Rows in fuel Excel files are checked first without writing anything: rows to add, rows already recorded and faulty rows are shown, quantities or unit prices above the limit are flagged row by row, and rows are added only after confirmation.

Document types

  • Live

    Vehicle registration

    Plate, chassis, engine, model year and owner details.

  • Pilot

    Policy

    Reading policy documents uploaded to the customer app.

  • Live

    Quote

    Reading insurers' price comparison documents in the panel.

  • Roadmap

    Claim files

    Accident reports, assessments, invoices and other claim documents.

What AI does · What you decide

What AI does

Reads the document, extracts fields, and adds a confidence score and location to each.

What you decide

You approve or correct the fields below the threshold. Your correction goes to the record as is and shows where the model struggles.

How it works

  • Field-level confidence

    The document is assessed field by field, not as a whole.

  • Boxes that show the source

    Each value is shown where it was read in the document, so checking takes one glance.

  • Review queue

    Uncertain fields collect in a queue, so no record is created with missing or wrong data.

Details

Technical details

Document types
Vehicle registration, price comparison PDF, fuel and EV-charging e-archive invoice, fuel Excel file. The insurer's own quote document is not read; it is fetched by quote number and previewed in the panel.
Where it is used
Quote creation in vehicle branches such as motor liability and motor own-damage, adding a vehicle on the customer card, publishing a quote request, fuel and charging records in the company-vehicle module.
Reading engine
A text-recognition model fine-tuned on Turkish registrations; runs on the CPU, no GPU needed. QR code first on newer registrations.
Where it runs
In-house service closed to the internet. Reached only through the signed-in endpoint of the panel or the mobile API.
Data handling
Registration images are processed in memory and discarded, never written to disk. The service has no write permission, and upload size is limited.
Status
Registration, price comparison and e-archive invoice reading are live in the panel. Registration scanning and policy PDF reading in the mobile apps are in pilot. Claim documents and layout-independent reading are on the roadmap.

Questions

Frequently asked questions

What does 85.7% measure?

The share of high-confidence values on the critical fields of real registrations that turned out correct. Exact-match accuracy across all fields is lower than this, and the real-document measurement rests on a small set. Higher precision measured on synthetic registrations does not carry over unchanged to real documents, so every value goes to a person for approval.

Are registration images stored?

No. An image is held in memory only while it is being read, and the service runs without permission to write to disk. The service does not keep the read values either; they reach the customer record only when staff save it.

Does the model act on its own?

No. A read value reaches the form only when staff press 'Apply to form'. In the platform design the model also only suggests: it never chooses which insurer to call, never produces or changes a premium, and never feeds an authorization decision. A person always starts binding actions such as issuing a policy, cancelling, making an endorsement or opening a claim.

Is it available in the mobile apps?

In pilot. In the staff mobile app, a registration can be read when adding a vehicle from the customer card, and the customer app is trialling registration scanning in its vehicle quote step. Reading a policy PDF from another agency in the customer app is also in pilot. That reader relies on label and value pairs, so it can still find the common labels when insurer layouts change, and it checks whether the policy belongs to the user or a registered family member. A scanned PDF without a text layer is reported as unreadable, and the details are requested by hand.

How much does photo quality matter?

A great deal; resolution comes before the model. On a low-resolution photo the model can produce a chassis number that is not on the document. The mobile apps therefore compress images at high quality so the small digits of the document serial stay readable, and the customer app does not send a photo below a minimum resolution.

Other parts

Back to platform