We start with managed services where specs are fluid and quality control matters — then productize the patterns that repeat.
| Lifecycle slice | Typical tasks | Worker tier | Quality controls | Pricing basis |
|---|---|---|---|---|
| Data acquisition | Speech, images, tool-use traces, domain documents | Generalists → specialists | Screening, consent, geo/language filters | Per item / project |
| Annotation & enrichment | Classification, extraction, transcription — with human verification of prelabels | Calibrated generalists + reviewers | Gold tasks, overlap, consensus | Per row / page / HIT |
| Alignment & post-training | SFT demos, pairwise ranking, preference data, reasoning traces | Specialists, vetted experts | Domain quals, rubrics, adjudication | Per judgment / SOW |
| Evaluation & safety | Benchmarks, factuality, hallucination, red team, model comparison | Experts, auditors, red teamers | Multi-stage QA, references, escalation | Pilot / retainer |
| Runtime oversight | Approvals, exception queues, agent escalation, tool-use intervention | Operators + domain experts | Audit logs, SLA routing, stop controls | Per queue / platform + usage |
Recurring capacity for model evaluation and assurance. Rubric-driven reviews, gold-set anchoring, weekly packs with scores, disagreements, and adjudicated outcomes.
Structured adversarial exploration with severity scoring and remediation notes — not unstructured “try to break it” freelancing.
Law, healthcare, STEM, coding, multilingual — when generalist crowd quality is the bottleneck.
Human approvals inside agent and document AI workflows. Reversible decisions, stop controls, and SLA-aware routing.
We do not position as an open marketplace for the cheapest labels. Routine first-pass annotation is increasingly software-assisted; our commercial center of gravity is decision-quality loops.
Data Day