Frequently asked questions
Getting started, quality, security and commercial terms. If something is missing, ask and we will add it.
Getting started
How quickly can a project start?
A pilot on an existing language and task type usually starts within a week. New languages, specialist domains or KYC-gated workforces take two to four weeks to recruit and qualify.
What do you need from us to quote?
Modality, volume, languages, any demographic or device constraints, your deadline, and ideally an example of what the model gets wrong today. Guidelines help but are not required to start.
Do you run a pilot?
On every project. A small paid batch against your real guidelines, which you accept before anything scales. It has saved every client who took it.
What is the minimum engagement?
There isn't a hard floor, but below roughly a week of production work the setup cost dominates and you are better served by a catalogue dataset.
Quality
What accuracy do you commit to?
A per-class threshold agreed during the pilot, not a single headline number. Averages hide the classes that matter.
How do you measure quality?
Hidden gold tasks in the live stream, two or three way consensus on ambiguous classes, a Bayesian quality score per contributor, and inter-annotator agreement reported with every batch.
What happens to rejected work?
It is reworked once at no charge where the rejection is in scope. Contributors can appeal, and appeals that succeed feed back into the guideline.
Do you report where annotators disagreed?
Yes, per class and per slice. If agreement is low we will tell you the metric is unreliable rather than shipping a number that looks clean.
Security and compliance
Where is data stored?
AWS Mumbai by default, with other regions available where residency rules require it. The region is written into the contract.
How is data secured?
AES-256 at rest, TLS 1.3 in transit, signed media URLs with no public access, role-based access with enforced two-factor authentication and full access logging.
Can we have a private workforce?
Yes. NDA-only pools, secure view mode with downloads disabled, and IP allowlisting on request.
What consent covers the data?
A consent record is attached to every file, naming the purpose and whether commercial use is covered. Withdrawal deletes the affected items.
Commercial
How does pricing work?
Per audio hour for speech, per image or per label for vision, per task for text and evaluation, or a monthly dedicated team. Catalogue datasets are priced on their own pages.
Who owns the data you deliver?
You do. Custom work transfers on full payment, including the contributor consents that make it usable commercially. Catalogue datasets are licensed and the terms are on each page.
Do you offer exclusivity?
On custom collection, yes. On catalogue datasets, exclusive licensing is available at a different price point.
Can we buy a sample first?
We send a representative slice including the hard cases, plus the QA report and licence terms, before any invoice.
Question not answered here?
Send it over. Someone from the delivery team replies, usually within a working day.