Verified, consented AI training data from South Asia and the Gulf.
A managed workforce captures speech, text, images, video and robotics demonstrations on ordinary phones. Every item is checked automatically, validated by consensus, paid per accepted unit and delivered with a datasheet and consent manifest.
- accepted items
- 104
- acceptance rate
- 95%
- languages covered
- 6
- live task types
- 23
- Text 68
- Image 32
- Robot lab 4
Counts read straight from the operational database on every page load.
- Faces and plates blurred before review
- Pay per accepted unit, no thresholds
How an item becomes training data
- 01
Capture
A worker records, photographs or writes on the phone they already own. Client-side checks catch problems before upload.
- 02
Automatic checks
A local Python worker runs speech, vision and metadata checks and attaches every score to the submission.
- 03
Human consensus
Two independent validators review what the machine cannot decide. Gold probes keep them honest.
- 04
Delivery
Accepted items enter a versioned dataset shipped with a datasheet and a per-item consent manifest.
What the platform collects
Urdu, Punjabi, Sindhi, Pashto, Saraiki, Balochi and Gulf Arabic dialects. Read, spontaneous and conversational.
Photo missions, captions and bounding boxes reviewed by consensus.
Chest-mounted phone capture with hand pose and motion tracks for physical AI.
Episode review on LeRobot-format datasets; teleoperation when hardware lands.