Computer Vision Engineer - Biometric Identification at Scale (1:N Open-Set Matching)

We're adding biometric identification to an established consumer mobile app. Users capture a subject with their phone camera at enrolment, and we later need to verify or identify that subject against our database. We built an initial prototype and quickly found this is a harder problem than it first appears. We're looking for someone who has genuinely solved 1:N identification before, not someone who has fine-tuned a classifier. The problem: - Enrolment and verification (1:1) from consumer phone cameras - uncooperative subjects, poor lighting, variable distance and angle - Identification (1:N) against a gallery of 100,000+ subjects, growing - Open-set: the correct answer is frequently "not in the database", so threshold calibration matters far more than top-1 accuracy - Significant visual near-duplicates within the population, plus template drift as subjects change over time Scope: - Feasibility phase: assess the approach, define realistic accuracy targets, specify the data volume and quality you'd need, and tell us plainly what won't work. We'd much rather hear "this can't hit your target at that gallery size" now than later. - Prototype: detection/alignment → embedding model → vector search, plus an evaluation harness we can run ourselves. We want FAR/FRR curves and per-cohort slices, not a single headline accuracy figure. - Production: on-device capture quality gating, server-side embedding and matching, indexing that holds up as the gallery grows. You should have: - Shipped a production 1:N biometric system — tell us which modality and at what gallery size - Deep metric learning experience (ArcFace/CosFace or equivalent) - ANN vector search at scale (pgvector, FAISS, Qdrant, Milvus) Mobile inference for capture-side quality checks (CoreML / TFLite) - Confidence building and defending an evaluation methodology To apply, answer these three (please skip the generic proposal): 1. What's the largest gallery you've run 1:N identification against, and how did accuracy hold at that size? 2. How would you set and validate the match threshold when most queries have no true match? 3. What dataset would you need from us before committing to accuracy targets?