More than a matching algorithm
A face-matching algorithm is one component of a verification system, not the whole thing. A production system also needs a liveness check ahead of matching, a defined confidence threshold, a path for near-tie results, and an audit trail for every decision.
The data problem
Building against a real, authoritative reference — a national registry photo, for instance — rather than a self-submitted photo is what makes verification meaningful rather than circular. That usually means integrating with an external registry, not just comparing two user-provided images.
The parts that get skipped early and regretted later
Rate limiting on face-verify endpoints, idempotency on enrollment calls, and a clear policy for how long captures and embeddings are retained are easy to deprioritize early and expensive to retrofit later.