datai.

Human actions.
Useful context.
Traceable data.

A recording becomes useful when its task, source, labels and permissions can be understood together. datai connects the people who capture physical activities with the teams that need that evidence.

Hands carrying out a precise manual task at a workbench
CAPTURE CONTEXT STARTS WITH THE TASK.

From a recording
to a licensed dataset.

Each stage keeps a clear boundary between what a contributor says, what software measures and what a model infers.

New automatic analysis is waiting for its production connection. Accounts, offers and private uploads are open; recordings stay private while processing is unavailable. Paid checkout and contributor payouts will open later.

01

Describe the activity

Contributors create a profile and publish an offer in a relevant activity category. They describe the work, choose a direct hourly price and decide whether their recordings can join category batches. Activity guides help turn a broad category into a specific, observable task.

02

Prepare the capture

Choose a viewpoint that shows the hands, objects and result. Confirm the right to supply the recording, participant permissions, workplace authorization and the removal of private information. Equipment arrangements are separate from marketplace registration. Start with the capture guide.

03

Upload the original privately

Submit the original MP4 or WebM with task context, environment, camera setup and permission declarations. Contributor action labels are optional. Upload completion does not make a recording public or available to buyers; its processing status remains visible in the contributor studio.

04

Measure the source

The media worker reads the original file, measures its technical properties and calculates a whole-file hash. These measurements remain distinguishable from contributor declarations and upload-part checksums. The original is preserved; processing derivatives and their settings are recorded separately.

05

Analyze and check

NVIDIA Nemotron analyzes silent video segments using sampled visual frames. It generates descriptions, approximate timed actions, objects, tools and quality or privacy flags. Versioned automatic gates check the complete package. Incomplete, uncertain or flagged recordings stay off sale; automatic acceptance is not a human review or proof of permission.

06

Freeze the licensed package

An order preserves its selected recordings, metadata, action labels and license version. Later uploads do not change that selection. Original downloads and frozen JSONL/CSV exports require a confirmed payment and a valid license when purchasing is available. The delivery reference explains the files and authenticated batch downloader.

Useful evidence.
Clear limits.

Acceptance makes a recording eligible for the marketplace. It does not guarantee a sale, scientific suitability or a successful robot behavior.

Keep the origin visible

Contributor context, measured media and generated analysis retain their sources and versions. A buyer can distinguish declared settings from verified measurements, and automatic decisions from exceptional human review. Data standards explain the baseline.

Scope additional signals

Sampled visual analysis does not inspect every frame or analyze audio. Ordinary video does not provide measured robot commands, calibrated depth, force or 3D pose. Masks, spatial tracks and other specialist labels need an agreed specification and suitable source material.

Define the permitted use

Contributors must obtain the necessary rights before recording and sharing. The marketplace’s non-exclusive license defines AI training and evaluation use; model privacy flags cannot establish legal permission. Read the license terms alongside the dataset requirements.

Start with the task.
Define the evidence.

Discuss your dataset