Skip to content
Telesurety

Research

Our research agenda

The platform's AI core rests on five open problems. They are too domain-specific to be solved by connecting to an off-the-shelf model, and each will be evaluated with its own measure.

Highlights

Registration reading baseline
85.7%
Field accuracy measured on real registrations; new models will be judged against this line.
Where models run
In-house
Personal data is processed only by models running on in-house servers in Türkiye.
How results are shared
Openly
Results will be shared on this page as written pieces, including experiments that fall short; the first pieces will appear once measurement results are ready.

Our only measured figure today is 85.7% field accuracy on real registrations, a single rate across all fields. Results by field type, and a separate measurement for document layouts never seen before, come next. Policy Q&A is at prototype stage; the other problems have no measured results yet.

Method

The measure first. Then the model.

  1. 01 / 05

    The test set is held out

    Measurement uses real documents never used in training. Insurer layouts the model has never seen are kept as a separate group.

  2. 02 / 05

    The baseline is written down

    The baseline is recorded before the work starts. For registration reading, today’s line is 85.7% field accuracy on real registrations.

  3. 03 / 05

    Models are tried

    Alongside today's registration reading model, layout analysis and small models with schema-bound output will be tried on the same test set; every field will come back with a confidence score.

  4. 04 / 05

    Measured field by field

    Results will be reported by field type, not as one average; unseen layouts will be measured separately.

  5. 05 / 05

    Corrections flow back

    Staff corrections are to be collected as labels. The aim is to reach the same accuracy with fewer labels, and to keep learning without losing accuracy on insurers already covered.

  6. Illustrative. Names and values shown are examples. The measure first. The test set is held out, The baseline is written down, Models are tried, Measured field by field, Corrections flow back.

Open problems

Five problems

  1. P1

    Field extraction from changing layouts

    Open question
    Can a small in-house model reach sufficient accuracy on an insurer's document layout it has never seen? How many labeled examples does a new insurer need?
    Approach
    Document-tuned text recognition and layout analysis; a small language or vision-language model with schema-constrained output; format and cross-check rules.
    How we'll measure
    Field-level accuracy, with unseen insurer layouts measured separately. Baseline: 85.7% on real vehicle registrations today.
  2. P2

    Confidence calibration

    Open question
    Does the model's confidence score match real accuracy? How should the automation threshold be set per field?
    Approach
    Calibration methods such as temperature scaling and conformal prediction; separate thresholds by field type.
    How we'll measure
    Calibration error and the error rate among fields above the threshold.
  3. P3

    Coverage ontology

    Open question
    Can different insurers' coverage phrases, including conditional coverages, be mapped correctly to one model?
    Approach
    A Turkish coverage ontology for motor own-damage, motor liability, home and earthquake; vector similarity with language-model re-ranking; expert approval.
    How we'll measure
    Mapping accuracy and inter-expert agreement.
  4. P4

    Question answering with citations

    Open question
    Can a small model stay faithful to policies and contracts, and learn to abstain when the documents don't support an answer?
    Approach
    Layout-aware chunking that keeps tables intact; mandatory citations in every answer; abstaining without support.
    How we'll measure
    Rate of correct, cited answers and rate of fabricated answers.
  5. P5

    Learning from corrections

    Open question
    How much can labeling effort be reduced by using staff corrections as labels?
    Approach
    Collecting corrections as labels, uncertainty sampling, and continual training that doesn't degrade performance on known insurers.
    How we'll measure
    Labels needed for the same accuracy, compared with random selection.

Research tracks

From reading the document to preparing the work for your approval.

Illustrative. Names and values shown are examples.

Document intelligence

Partly live

Document intelligence that starts with registrations.

The AI running in production today is registration reading: a text-recognition model fine-tuned on Turkish vehicle registrations and run in-house. It measured 85.7% field accuracy on real registrations, and that figure is the research baseline. Extending the same method to policies, quotes, driving licences, IDs, invoices and claim documents is on the roadmap.

  • Every value passes two checks: a fixed format rule and a model-confidence threshold set from measurement. If either is missing, the field is flagged for review.
  • On newer registrations, the document's own QR code is decoded first. In tests on real registrations, it corrected serial letters the text reader had returned wrongly with high confidence.
  • Images are processed in memory and discarded, never written to disk. A value reaches the form only when the user transfers it.
  • Confidence calibration is a research topic of its own: on real registrations, raising the threshold does not yet buy higher precision.
  • Roadmap: layout analysis will work with a small in-house language or vision-language model fine-tuned with LoRA. Output will be held to a fixed schema, and each field will return its value, confidence score, location on the document and model version.
Illustrative. Names and values shown are examples.

Question answering with sources

Prototype

Can a small model stay faithful to the document?

What runs today is a prototype: it answers Turkish questions about an uploaded policy or quote PDF, shows the page number behind each document-based answer, and runs entirely in-house. Contract intelligence is on the roadmap: over agency and web service agreements and data protection notices, it will answer with the clause each answer rests on, and say there is no provision when the documents give no basis.

  • The prototype runs on a small open-weight language model, a local embedding model and a self-hosted vector database. It calls no outside AI service.
  • For coverage-limit questions, when the document has amount lines, the language model does not write the answer; coverage lines and amounts are taken directly from the document.
  • Its limits are clear: citations are page-level, with no clause-level citation yet, and it cannot yet decline when the context falls short. Teaching the model to stay silent is the heart of the research.
  • Method, on the roadmap: layout-aware chunking that keeps tables intact, a mandatory citation in every answer, and no answer without support.
  • Measurement, on the roadmap: the rate of cited correct answers and the rate of fabricated answers, measured separately. Targets have been written as proposals; there are no measured results yet.
Illustrative. Names and values shown are examples.

AI Assistant and Agents

Roadmap

An in-house model that will know insurance, and agents that will prepare the work for your approval.

The assistant will run on an in-house language model of at least 12B parameters. It will not be trained from scratch: it will be built on an open-weight model and adapted with LoRA to insurance terms, schema-conformant output and quoting its sources. Facts will always come from documents and records. All of this is on the roadmap; no model of this size runs today.

  • The policy and regulation assistant will answer by showing the relevant page; if it finds no basis, it will say it could not find one.
  • The quote and renewal agent and the inbox and claims agent will use the panel's own endpoints as tools, with your permissions; no separate business logic will be written for them.
  • Insurer selection, premiums and permissions will stay with rules. The model will not estimate prices, and every irreversible action will wait on an approval card.
  • Turkish voice commands are a separate research topic: converting Turkish speech with insurance terms and plates to text in-house. Typing will not wait for that research.
  • The foundation running today: the staff app prototype, the policy Q&A prototype and the panel's live server-side permission checks. No model version will go live before passing test sets prepared by experts.

Infrastructure and principles

Where the research runs, and by what rules.

Where models run
In-house. Registration reading runs on the CPU with no GPU, is closed to the internet and writes nothing to disk. Planned models will also run on in-house servers in Türkiye; no model will run on the phone.
Training data
Personal data is processed only by in-house models. An outside model may act only as a 'teacher', generating training examples from anonymous or synthetic data. An organization's data will be used for training only with its explicit consent and after anonymization.
Release rule
A model version that lowers the touchless record rate on the evaluation sets will not be released.
Today's baseline
85.7% field accuracy on real registrations. Today's rule-based readers, and off-the-shelf text recognition with a general language model that has not been fine-tuned, will also serve as baselines for comparison.

Questions

Questions about the research

Which part of the research is in use today?

Registration reading is live, in the operations panel's quote and customer forms. Question answering with sources over policy and quote PDFs is a prototype. Layout-independent reading and claim documents, contract intelligence, the Turkish coverage model and the AI assistant are on the roadmap.

How was 85.7% measured?

As field accuracy on real registrations, and the measurement rests on a small set of real documents. Higher precision measured on synthetic registrations does not carry over unchanged to real documents. That is why every value goes to a person for approval, and why new models are measured on real documents.

Why not plug in an off-the-shelf large model?

The problems are too domain-specific to be solved that way, and personal data is never sent to an AI service outside the company. Instead, small open-weight models will be adapted in-house with LoRA to insurance terms, schema-conformant output and quoting their sources; no model will be trained from scratch. A large cloud model may act only as a 'teacher', working on anonymous or synthetic documents.

What happens when the model is unsure?

Every field comes with a confidence score. A field below the threshold, or one that fails its format rule, is never written to the record on its own; it goes to a person for approval. Whether the model's confidence matches its real accuracy is a research problem of its own: confidence calibration.

Will our corrections be used to train models?

As a principle, staff corrections will become training data only with the tenant's explicit consent and after anonymization. The aim is to reach the same accuracy with fewer labels, and to keep learning without losing accuracy on insurers already covered.

Where will the results appear?

In the writing on this page. Nothing has been published yet; the first pieces will appear here once measurement results are ready. Experiments that fall short will be shared as well.

Results

Writing

Nothing published yet. The first pieces will appear here once measurement results are ready.