Intelligent assisted onboarding
A deposit-account application moves through consent, identity and document facts, and a next action. The workflow, not a model, decides what happens next, and it can always say why.
Real: the engine and its tests
- Deterministic workflow engine with explicit consent blocking.
- Facts carry a source citation and a confidence score; missing or low-confidence facts route to review.
- Market rules are isolated per country (US and India policies in the repository).
- Core write-back is idempotent: a repeat returns the original result.
- Ordered audit events for every step. 52 executable tests pass on the current code (covering consent, KYC gating, provenance, missing fields, resume, idempotency, market isolation and audit order).
Synthetic: the data and the partners
- Applicants come from four generated fixtures.
- The ID check score and the core-banking write are simulated.
Scale check
10,000synthetic applicants run9,108completed with a core write892routed to review or decline handling
Fixed seed 20260915 with the repository command npm run cohort; rerunning gives the same counts. Of the 10,000, the simulated KYC check passed 9,366, sent 409 to manual review and declined 225. The generated applicants are all in the India market, with consent given and every extracted fact at full confidence, so the outcomes follow labels assigned in the generator. These are deterministic synthetic results, not business metrics. Live demonstration access is by request.