← All work

RavoTeq, for an identity verification company · 2026 to now

Identity verification platform

Identity and business documents read, labelled and checked in real time during onboarding, so a reviewer signs off on evidence instead of re-reading every file. 10K+ users served globally.

10K+users served globally

The problem

KYC onboarding stalls on reading documents. Someone has to open every passport, certificate of incorporation and utility bill, work out what it is and check it against the rest of the file. And the documents are too sensitive to send to a general-purpose AI service.

Participants

+ Add participant

12Pending to be sent

38Ongoing screenings

214Positive one-time

96Positive ongoing

NameScreeningStatus

Sanne Visser0172836287Basic ScreeningSUCCESS

Daan Mulder0172836311Safety ScreeningACTIVE

Priya Raman0172836402Basic ScreeningNOT STARTED

Lucas de Graaf0172836455Safety ScreeningSUCCESS

Emma Bakker0172836498Basic ScreeningNEGATIVE

Noah Jansen0172836520Basic ScreeningACTIVE

My role

CTO at RavoTeq, engineering, architecture and the document AI

Results

Real time
document labelling inside the onboarding flow

Stack

  • PyTorch
  • vLLM
  • Pydantic AI
  • DeepEval
  • Custom SLM

What changes

Document review

WithoutReviewers open every passport, certificate and bill to work out what it is.

With Identity verification platformDocuments arrive labelled and classified in real time, inside the onboarding flow.

Sensitive data

WithoutIdentity documents sent to a general-purpose AI service leave your control.

With Identity verification platformInference runs on a small model, on infrastructure we control.

Reviewer trust

WithoutAutomated results a compliance reviewer cannot explain.

With Identity verification platformEvery result reaches the reviewer with its evidence. A person makes the final call.

Change risk

WithoutA model update can quietly make results worse.

With Identity verification platformEvery change is evaluated offline and online before it reaches a customer.

How it works

At a high level. A detailed walkthrough is available on a call.

  1. 01Capture
  2. 02Classify
  3. 03Serve
  4. 04Orchestrate
  5. 05Evaluate
  6. 06Review
  1. Decision 01

    A fine-tuned small language model instead of a general-purpose API for document inference.

    Sensitive documents stay on infrastructure we control, cost per document stays predictable and latency fits inside the onboarding flow.

  2. Decision 02

    Typed state and tool contracts for every agent.

    A malformed result fails loudly at the boundary instead of quietly reaching a reviewer.

  3. Decision 03

    Evaluation built in from the first release, offline and online.

    Every change is measured before it ships, which is what a compliance team needs to trust an automated check.

  4. Decision 04

    The model labels and flags. A person approves.

    Automation removes the reading, not the accountability.