Design Face Recognition - System Design | WinJob

Biometric face platform at Apple/Google/Amazon scale

Design a face recognition platform that supports 1:1 verification (is this the same person?), 1:N identification (who is this among millions enrolled?), and watchlist screening with strict privacy controls. Apple interviews stress on-device templates and liveness; Google Photos stresses clustering at billion-image scale; Amazon Rekognition stresses multi-tenant galleries and auditability.

Personas

Mobile unlock: on-device embedding, Secure Enclave, no raw face leaves phone
Enterprise KYC: cloud verify against government ID portrait with liveness
Physical security: camera NVR streams → edge box → central gallery match
Consumer photo app: unsupervised face grouping with user consent toggles

Core pipeline every interviewer expects

Detect face bounding boxes (MTCNN/RetinaFace)
Align to canonical pose (similarity transform on landmarks)
Embed 512-d ArcFace-style vector
Score cosine similarity vs threshold calibrated to FAR/FRR
Policy: consent ledger, retention TTL, jurisdiction flags

Scale anchors (state aloud)

800M 1:1 verifications/month (~310 avg QPS, 12k peak)
120M enrolled identities across tenant galleries
40k 1:N search QPS peak for access-control customers
Target FAR 1e-6 at FRR 0.3% for banking-grade verify

javaOne Dark Pro

1 public record FacePlatformContext(String tenantId, String subjectId, String galleryId, double matchScore) {}

pythonOne Dark Pro

1 @dataclass(frozen=True)
2 class FacePlatformContext:
3     tenant_id: str
4     subject_id: str
5     gallery_id: str
6     match_score: float

typescriptOne Dark Pro

1 export interface FacePlatformContext {
2   tenantId: string;
3   subjectId: string;
4   galleryId: string;
5   matchScore: number;
6 }

Why interviewers care

Face Recognition interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

Interview checkpoint

Name one failure story for Problem Statement: Enterprise Face Recognition Platform that proves you understand real outages, not happy-path diagrams.

Key Highlights

•Anchor sec-01 on FAR/FRR and gallery sharding
•Keep policy engine outside the neural graph
•Quote 12k peak verify QPS before GPU SKU selection

1	@dataclass(frozen=True)
2	class FacePlatformContext:
3	tenant_id: str
4	subject_id: str
5	gallery_id: str
6	match_score: float

1	export interface FacePlatformContext {
2	tenantId: string;
3	subjectId: string;
4	galleryId: string;
5	matchScore: number;
6	}