Product Context
LabelFlow is a managed data labeling platform used by enterprise ML teams to annotate text, images, audio, video, and document/PDF data. Customers upload tasks, and the platform must route each item to the right annotator queue, pre-label model, and quality-control workflow while keeping turnaround time and cost low.
Scale
| Signal | Value |
|---|
| Enterprise customers | 4,000 |
| Daily active annotators | 120,000 |
| Tasks created per day | 45M |
| Peak task-ingest QPS | 18K |
| Peak labeling-assignment QPS | 35K |
| Active task backlog | 600M items |
| Input modalities | text, image, audio, video, PDF/document |
| p99 latency budget for assignment API | 250ms |
Task
Design an end-to-end ML system for a multi-input-type labeling platform. Your design should address:
- How to represent heterogeneous inputs and build a multi-stage decision system for task routing, annotator matching, and optional pre-label generation.
- The online and offline architecture, including feature stores, model training, batch vs real-time inference, and feedback logging.
- How to support cold-start for new customers, new task types, and new annotators while meeting latency and cost constraints.
- How to evaluate the system offline and online across quality, throughput, cost, and fairness.
- The main failure modes, especially feature drift, training-serving skew, and quality regressions by modality or customer segment.
Constraints
- Some customers require data residency and cannot send raw data across regions.
- Human-in-the-loop quality is the product: wrong routing increases rework cost and SLA misses.
- Raw video/audio inference is expensive; heavy multimodal models cannot run synchronously on every request.
- New task schemas appear weekly, and label taxonomies can differ by customer.
- The platform must provide auditable assignment decisions for enterprise compliance.