How we work

Data quality is a matter of mechanism, not assurance

The real risk in outsourced data is receiving something that looks like data but cannot be used — and only discovering it after the model has trained. This page states exactly what stops that.

From brief to delivered data

Agree brief and quote → Collect or ingest → Annotate and verify → Deliver and reconcile → Below bar → redone01Agree brief andquote02Collect or ingest03Annotate andverify04Deliver andreconcileBelow bar → redone
  1. 01

    Agree the brief

    We pin down data type, volume, quality bar and delivery format. Your quote itemises unit prices per line with concrete dates — not one blended number and a promise of 'as soon as possible'.

  2. 02

    Collect or ingest

    For collection work, tasks go out to contributors with location and content constraints. For annotation work, you upload through the dashboard or API. Both paths feed the same processing queue.

  3. 03

    Annotate and verify

    AI pre-labels, humans verify and correct, then work passes through peer verification and gold-standard checks. Anything that fails a layer goes back to be redone rather than through to delivery.

  4. 04

    Deliver and reconcile

    You receive the agreed format with a quality report for the batch. Anything below the committed accuracy is redone at no charge — that sits in the contract.

Four control layers on every submission

These four run automatically, independent of whether anyone remembered to check. The diagram describes the filtering mechanism, not the actual rejection rate of any specific batch.

Submissions in → AI pre-check → Fraud prevention → Peer verification → Gold standard → Approved and paidSUBMISSIONS IN01AI pre-check02Fraud prevention03Peer verification04Gold standardApproved and paid
01

AI pre-check at the door

Every submission is checked for format and matched against the task brief before it enters the queue. Failures are blocked immediately and cost no reviewer time. Where the machine cannot conclude, the case goes to a human — it does not default to pass.

02

Fraud prevention from data, not judgement

Content hashing blocks duplicate submissions, GPS coordinates plus a radius rule prevent repeat farming of one location, and rate limits stop automation. All three are hard rules, not a reviewer's subjective call.

03

Peer verification

Cases where the AI is not confident go to multiple independent reviewers. Only consensus passes — one person saying yes is not enough for a submission to get paid.

04

Continuous gold standard

Known-answer samples are mixed into live work to measure each contributor's accuracy over time. Anyone who drops below the threshold is restricted — onboarding tests are not enough, because a person's quality changes month to month.

Why a cheaper vendor usually costs more

Splitting the work between two vendors — one to capture, one to annotate — looks cheaper on the quote. But nobody signs for the gap between them, and you are the last to find it.

Annotation vendor / Collection vendor / TucTakCAPTUREANNOTATEDELIVERAnnotation vendor—✓✓Collection vendor✓——The gap you absorbTucTakCaptureAnnotateDeliver

What sits in the contract

Redone at no charge

Redone at no charge

Anything below the committed accuracy is redone free. This is a clause, not goodwill.

Explicit ownership

Explicit ownership

Data you send us stays yours. Data we collect under your brief is yours as well.

NDA before any file

NDA before any file

A confidentiality agreement is signed before any of your files enter our systems.

Concrete dates

Concrete dates

Quotes carry a delivery date, not 'depends on volume'. If we are going to be late, you hear it in advance.

Prefer to verify rather than read?

Send a small batch. Real quality is visible after a few thousand images, faster than any presentation.

Send a pilot batch