Ground truth, gathered on the ground.
Scraped data plateaus and synthetic data drifts. Hannu sends verified people to collect what the internet cannot see — prices, conditions, responses, labels — against your schema, with proof attached to every row.
Schema-validated rows · Gold-task quality gates · Consent recorded · One signed dataset
How it runs, end to end.
You define the schema
Fields, locations, sample size, the proof each row needs. Escrow locks the full run before collection starts.
Verified Operators collect
Matched by zone and skill. Hidden gold tasks with known answers run inside the batch — quality is measured, not assumed.
Rows verified, dataset signed
Each submission passes GPS, freshness and content checks plus human review where flagged. You receive one dataset with a recomputable audit trail — and pay only for rows that pass.
Data collection
On-the-ground gathering against a defined schema — markets, outlets, streets.
Surveys
Structured surveys with real, consented respondents in the field.
Data labeling
Human labels for training and evaluation sets, gold-task quality-gated.
Multi-site market survey
A structured study across many sites, returned as one signed workflow deliverable.
How your agent calls it.
Declare the schema once; every submission is validated against it before it ever reaches you. Batch endpoints handle multi-row runs.
# from your agent, via the hannu MCP server
create_task({
type: "data_collection",
zone: "yaba-lagos",
brief: "Record price + stock for 12 SKUs at open-air markets",
schema: { fields: ["sku", "price_ngn", "in_stock", "photo"] }
})Asked before every first task.
How do you keep collectors honest?
Three layers: KYC-verified identity (no anonymous resubmission under new accounts), hidden gold tasks with known answers scored inside every batch, and AI checks — GPS, EXIF, freshness, duplicates — with human review on anything uncertain.
Is respondent consent handled?
Yes — survey templates capture respondent consent as part of the proof, and Operator identity is never exposed to respondents. Data handling follows the NDPA; see the trust page for the full posture.
What formats do I get back?
A structured dataset matching your declared schema, with per-row proof references and a signed Verified report for the run — JSON via the API, or export from the console.
Can I run this at training-data scale?
The pilot runs one Lagos zone with a deliberately small verified cohort — real capacity, honestly stated. Volume runs are a committed-volume conversation; the multi-site workflow unlocks per zone as gate metrics hold.
“The scarcest training data is the kind someone actually walked outside to collect.”