Design Image Recognition - System Design | WinJob

Enterprise image recognition platform

Design a multi-tenant visual intelligence API comparable to Google Cloud Vision, Amazon Rekognition, or Clarifai: developers send photos or URLs, receive labels, bounding boxes, OCR text, embeddings, or custom classifier scores.

Interview framing

Google and Amazon interviews probe whether you treat vision as distributed systems + ML ops, not a single ResNet in a notebook. Clarifai-style products add custom model hosting and vector search.

Core capabilities to name early

Synchronous classify/detect/OCR with strict p99 latency
Async batch jobs over millions of images in object storage
Custom model upload, versioning, canary promotion, rollback
Embeddings + similarity search for dedup and visual search
Safety: NSFW/violence classifiers, PII blur, geo restrictions

Scale anchors (state explicitly)

2,000,000,000 images/day → ~23,148 avg QPS, 185,000 peak (8×)
Average 450 KB JPEG after client compression; heavy 4K bursts for enterprise tenants
8,000 custom models; 1200+ GPU nodes across 6 regions

Architecture split interviewers expect

Control plane: model registry, routing table, quota/billing, policy config. Data plane: preprocess workers (decode, resize, EXIF strip), GPU inference (Triton/TensorRT), post-process (NMS, thresholding), optional vector index write.

Failure modes to volunteer

Poisoned uploads (decompression bombs), EXIF malware, adversarial patches, hot keys on celebrity detection models, GPU OOM from greedy dynamic batching, embedding index lag after model swap.

javaOne Dark Pro

1 public record VisionRequest(String tenantId, String imageUri, String modelId, Map<String,String> opts) {}

pythonOne Dark Pro

1 @dataclass(frozen=True)
2 class VisionRequest:
3     tenant_id: str
4     image_uri: str
5     model_id: str
6     opts: dict[str, str]

typescriptOne Dark Pro

1 export interface VisionRequest {
2   tenantId: string;
3   imageUri: string;
4   modelId: string;
5   opts: Record<string, string>;
6 }

Why interviewers care

Image 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 Image Recognition Platform that proves you understand real outages, not happy-path diagrams.

Key Highlights

•Separate control plane (registry, routes, policy) from data plane (preprocess + GPU)
•Quote 2B images/day before naming GPU counts
•Offer sync API plus async batch over object storage

1	@dataclass(frozen=True)
2	class VisionRequest:
3	tenant_id: str
4	image_uri: str
5	model_id: str
6	opts: dict[str, str]

1	export interface VisionRequest {
2	tenantId: string;
3	imageUri: string;
4	modelId: string;
5	opts: Record<string, string>;
6	}