Design NLP Pipeline - System Design | WinJob

Problem Statement: Enterprise NLP Processing Platform

Enterprise NLP at Hyperscale

You are designing a multi-tenant NLP platform that ingests unstructured text (support tickets, product reviews, news, contracts) and returns structured signals: named entities, sentiment, intent, and topic labels. Google, Amazon, and Microsoft ask this to test whether you understand batch vs real-time, model serving, and data governance—not a single REST endpoint calling an API.

Personas

Product engineer calls POST /v1/analyze with 2 KB snippets needing <200 ms p99.
Data platform team runs nightly batch jobs over 40 TB corpora with cost-optimized GPU pools.
ML owner ships new ONNX exports weekly with canary promotion and automatic rollback on F1 regression.

Core journeys

Upload document → job id → poll or webhook → enriched JSON with spans and scores.
Synchronous analyze for chat widgets with strict tenant isolation and PII masking.
Reprocess corpus when model v3 improves NER for a new entity type.

Why this is hard

Language detection errors cascade; entity resolution merges "MSFT" and "Microsoft"; GPU memory limits batch sizes; poison documents (binary masquerading as text) stall workers; regulators require redaction before persistence.

Interview checkpoints (1/30)

Tie decisions to measurable NER F1, p99 latency, and $/1k documents
Call out tenant isolation whenever persistence appears

javaOne Dark Pro

1 public record AnalyzeRequest(String tenantId, String text, List<String> tasks) {}

pythonOne Dark Pro

1 @dataclass(frozen=True)
2 class AnalyzeRequest:
3     tenant_id: str
4     text: str
5     tasks: list[str]

typescriptOne Dark Pro

1 export interface AnalyzeRequest {
2   tenantId: string;
3   text: string;
4   tasks: string[];
5 }

Why interviewers care

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

Interview checkpoint

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

Key Highlights

•Anchor Problem Statement: Enterprise NLP Processing Platform to measurable NLP SLOs
•Keep tenant_id on every persistence and cache key

1	@dataclass(frozen=True)
2	class AnalyzeRequest:
3	tenant_id: str
4	text: str
5	tasks: list[str]

1	export interface AnalyzeRequest {
2	tenantId: string;
3	text: string;
4	tasks: string[];
5	}