Instant biometric onboarding

The context

A fintech acquiring customers through field agents, far from branch infrastructure. A new account began as paper forms and photocopied identity documents, then took about two days to reach a back office before the customer existed in any system. Field connectivity was unreliable, so cloud-only tools were ruled out early. The constraints were fixed: identity verification had to be trustworthy, fraud caught at capture, and the whole flow run on a device in the agent's hand.

Why normal software was not enough

A digital form alone would have moved the paperwork, not the judgment. The slow part was verification: does this face match this document, is this document genuine, is this person live and present. That is perception, and rules cannot do it. Face matching, document reading and liveness checks needed models, and they had to run on the device because the connection could not be trusted to exist.

The problem

Field onboarding meant paper forms, photocopied identity documents, and a two-day round trip to a back office before a customer existed in the system. Connectivity in the field was unreliable, so a cloud-only solution was never going to survive contact with reality.

What we built

We built an onboarding flow around on-device intelligence: face recognition matched the applicant to their identity document, document analysis extracted and validated the fields, and liveness checks stopped photo-of-a-photo fraud. Everything ran offline first and synced when a connection appeared, with human review queues for the cases the models flagged as uncertain.

What production made hard

• Offline first is an architecture, not a feature flag: capture, matching and validation all had to complete on the device, with sync designed around connections that appear and disappear.
• Field-captured documents are messy: glare, worn cards, low light and camera shake degrade extraction, so every field needed validation rules and confidence scores rather than blind trust.
• Liveness checks had to stop photo-of-a-photo fraud on ordinary phone hardware without frustrating legitimate applicants, which meant careful tuning between fraud caught and good customers blocked.
• On-device inference forced real trade-offs: model size against accuracy, speed against battery, on the devices agents actually carry.
• The review queue was its own design problem: deciding which uncertain matches route to humans, and what the agent tells the customer while a case waits.

What a similar project needs

• Data: identity documents and consented face imagery captured in real field conditions, not clean studio samples. If only clean samples exist, budget time to collect messy ones.
• Sequence: a prototype sprint can prove the core capture-and-match flow; the larger effort is hardening for offline sync, fraud cases and device coverage.
• Team: computer vision and mobile engineering working as one unit, because the models and the device constraints have to be designed together.
• Risks: biometric data demands privacy and consent handling from day one, with PII masking, least-privilege access and audit logs designed to support DPDP and GDPR expectations.
• A human review path for low-confidence matches, planned before launch, so uncertain cases become queue items instead of silent failures.