About The Dataa
We started because the data we needed did not exist
The Dataa was founded by people who spent years trying to ship speech and language products to Pakistani users, and kept hitting the same wall: the training data was not there, and the vendors who claimed to have it did not.
Our thesis
The next order of magnitude in model quality is not compute
It is coverage. The frontier of what models can do in English is genuinely extraordinary. The frontier of what they can do in Saraiki, Balochi or Brahui is roughly where English was in 2015 — and 255 million people live on that frontier.
The gap is not a modeling problem. Architectures transfer. Scaling laws transfer. What does not transfer is data that was never collected in the first place. There is no Common Crawl of spoken Sindhi. There is no Reddit of Saraiki medical advice. There is no archive of how a Pashto speaker actually asks a bank for help.
Somebody has to go and get it. That requires field teams, local trust, consent frameworks that work at low literacy, payment rails for unbanked contributors, and a quality system that catches fraud at scale. It is unglamorous, operationally heavy work, and it is the entire reason we exist.
Our position on the ethics: we think this work is only defensible if the people producing the data are paid properly, understand what they are consenting to, and can walk away. We also think that is not a trade-off against quality — it is the precondition for it. Every shortcut on consent or pay shows up later as fraud, low variance and re-collection cost.
Principles
Five things we will not trade away
These have cost us deals. We are writing them down anyway, so you can hold us to them.
We tell you when the answer is no
Some languages have too few literate speakers to staff a QA team. Some districts we will not send collectors into. Some specs will not produce the outcome the client wants. You will hear that before you build a roadmap on it, even when saying so costs us the work.
Contributors are paid properly, and know why
At or above local living wage, disclosed in writing beforehand, paid within 14 days, through channels that work for people without bank accounts. What we pay contributors, by role and province, is disclosed to clients on request.
Consent is real or the data does not ship
In-language, at a tested reading level, explicit about model training, with a working withdrawal channel and an honest statement of what withdrawal cannot undo.
We report the numbers that make us look bad
Inter-annotator disagreement, rejected batches, quota shortfalls and missed milestones go in the delivery report. A vendor who only reports good news is a vendor you cannot calibrate.
Small deliveries early, always
You see real data in week two, before we scale headcount. Course corrections are cheap early and ruinous at final delivery.
Local teams, not remote extraction
Every office is staffed and led by people from that province, with hiring, pay and career progression in-country. Karachi is head office, not a coordination desk reporting somewhere else.
Leadership
Who runs this
A deliberately unglamorous mix: speech researchers, operations people who have run large field workforces, and a compliance lead who used to review vendors like us.
Leadership is not listed yet. We would rather show nothing here than invent a team. If you are evaluating us and want to know exactly who would own your project, who signs off on consent, and who you escalate to when something slips, ask and we will introduce them directly.
Careers
We hire in the provinces we collect in
Every one of our staff is based in Pakistan, in permanent roles with local career progression — not contract coordination seats.
Applications from native speakers of under-served languages are especially welcome, particularly Balochi, Brahui, Shina and Khowar.
No formal openings right now
We hire in bursts, when a project lands. If you are a native speaker of an under-served language — particularly Balochi, Brahui, Shina or Khowar — or you have run field collection, QA or annotation operations before, send your CV anyway. We keep them and we do come back to them.
Press & research
We publish what we learn
Methodology notes, dataset datasheets and the occasional uncomfortable finding about how this industry actually operates. Media enquiries get a same-day response.
- Dataset datasheets published for every catalog corpus
- Methodology write-ups on quota sampling and consent operations
- Transfer studies: measured WER deltas when Urdu data is used to improve Hindi ASR
- How data-collection labour is actually paid across South Asia
- Media kit, logos and executive bios available on request
Come and see the operation.
We host client visits at our Karachi, Lahore and Islamabad offices. Watching a collection session run is more persuasive than anything on this website.
Average first response: under 6 business hours. NDAs signed same day.