Six signal groups resolve one application against its credentials, device, economic footprint and connected infrastructure.
Production modules use the same six signal definitions and evidence language.
Compares names, dates, addresses and application origin across credentials.
Reads issuance age, linkage, CKYC history and identifier reuse.
Checks mobile tenure, recent telecom events, device reuse and location alignment.
Separates a thin credit file from an absent or manufactured pattern of activity.
Maps shared addresses, beneficiary accounts, face matches and submission timing.
Measures how identity fields were entered against the form baseline.
A buyer should not have to join a consortium to learn whether the product works on its own book.
Identity consistency, credential vintage, telecom and device evidence, economic footprint and behavioural signals work against the institution's own application data.
Address, device, beneficiary and face reuse become stronger as prior applications and outcomes are resolved into the institution's graph.
Permitted consortium matches can reveal a ring that crosses institutions without exposing another member's customer record.
A member receives standalone value before any cross-institution signal is enabled. Network evidence is an additional source, not a requirement for a score.
Entity resolution turns tokenised evidence into an identity graph. A graph model evaluates node and relationship evidence, while segment calibration maps the result to policy bands.
Hard overrides cap the score when a revoked mobile, repeated face vector or high-velocity device makes the graph evidence decisive.
Most banks will not send onboarding records to a vendor cloud. The scoring interface does not require them to.
Tokenised requests reach an India-region service boundary. Raw direct identifiers do not leave the member environment.
The inference service runs in an isolated network boundary connected to the institution's onboarding systems.
Banks that cannot send onboarding evidence to a vendor cloud can run the scoring service inside their own environment.
Historical files are scored in a controlled batch environment before any production integration is approved.
The synchronous surface stays small. Evidence expansion and case review use the application identifier returned with the decision.
{
"applicationId": "app_20481",
"identityToken": "tok_7fd2",
"mobileAgeDays": 728,
"bureauHistoryMonths": 0,
"addressReuseCount": 1
}{
"score": 64,
"band": "Thin",
"recommendedAction": "step_up",
"reasonCodes": [
"thin_footprint",
"identity_coherent"
]
}The feedback webhook records confirmed fraud, clean performance and analyst outcomes. Federated training can then improve local weights without moving the member's customer rows into a common table.
A model should improve from the institution's labels without taking ownership of its customer data.
We do not publish an unattested response-time claim. The pilot measures service processing and network transit under the institution's expected load, then writes the production SLO against that result.
The retrospective backtest exercises the signal model and produces the first calibration discussion before production architecture is chosen.
Book a backtest