/ Labs · Interview continuity
Hiring fraud · candidate substitution

Is the person in round 4 the same human from round 1?

Each interview round captures a face embedding in the browser and derives a fuzzy person-key from it — no server, no account, no login. The key is computed from fixed public hyperplanes, so the same face produces a near-identical key on any device, in any session. A stand-in candidate produces a key that disagrees in far too many bits to be the same person.

Privacy posture: the raw face image never leaves the camera element. The 128-float embedding and the derived 64-bit key are stored only in this browser's localStorage. Nothing is uploaded. The person-key is one-way — you cannot reconstruct a face from it.
Tip: capture two rounds yourself, then have someone else do round 3 — watch the key diverge. Sessions persist across reloads (and the key math is device-independent).
How it works: face-api.js computes a 128-dim FaceNet descriptor → the descriptor is projected onto 64 fixed, seeded random hyperplanes → the sign of each projection becomes one bit of the person-key (locality-sensitive hashing). Same person ⇒ small Hamming distance (typically ≤ 10 of 64 bits). Different person ⇒ ~32 bits, indistinguishable from random. Matching thresholds: euclidean < 0.6 on raw embeddings, Hamming ≤ 16 on keys. This is probabilistic evidence, not cryptographic proof — combine with liveness checks and a device-bound passkey for the full Manav receipt.