How it works

From integration to network detection, in four steps

YARDA trains its models on your transactions, without names or addresses, with pseudonymised identifiers. They learn from your business and improve as your teams label alerts: detection built on your reality.

  1. Step 1 · Integration

    Connect your transactions

    Your system sends each transaction to our API, or your team uploads a file from the console. Identifiers are pseudonymised as soon as they arrive, and names and addresses are never kept. YARDA relies on the characteristics of the transaction, such as the amount, currency, corridor and timestamp.

  2. Step 2 · Anomaly detection

    Detect anomalies

    The first model learns how your business normally behaves from your transactions. It needs no prior labels and, from the day it goes live, flags the transactions that depart from that behaviour: the ones worth a closer look.

  3. Step 3 · Supervised classification

    Classify fraud

    Your teams review the alerts and label each transaction with your own fraud taxonomy; a single transaction can fall under several types. These labels train a supervised model that learns to recognise each type of fraud specific to your business, and that improves as labels build up.

  4. Step 4 · Network detection

    Uncover networks

    Organised fraud does not show one transaction at a time. Every day, YARDA maps the links between your customers: who pays whom, who shares a device. The console draws networks from them, such as accounts that received confirmed fraud paying the same recipient, several accounts used from one device, or money going round in a cycle, and shows your analysts the links around a customer, an alert or a case. A network is a lead to examine, not proof. A graph neural network is being researched; it plays no part in scoring.

API integration

A simple API: one transaction, one score, one explanation

Send the details of a transaction and receive a risk score, a decision and the factors behind it. A documented REST API, usable from any system that can send an HTTPS request.

POST /v1/predict/fraud
{
  "transaction": {
    "transaction_id": "TRX-2026-000184",
    "timestamp": "2026-10-18T09:42:11Z",
    "amount": 150000,
    "currency": "XOF",
    "sender_id": "CLI-004512",
    "receiver_id": "CLI-019877"
  }
}
Response 200 OK
{
  "decision": "alert",
  "final_score": 0.78,
  "decision_cause": {
    "kind": "score_band",
    "reason_fr": "Score de 78, au-dessus du seuil d'alerte (70)."
  }
}

See YARDA on your own transactions

Tell us who you are: we will get back to you to arrange a demo, then a free diagnostic on your history. No development, no commitment.

What happens next

  1. A demo of the console, on fictitious data
  2. A free diagnostic on three to six months of your history
  3. A pilot with success criteria agreed together