There are two ways to get data from Syntropi. License footage immediately from our existing library of over a million cleared videos, or commission a custom assignment to capture exactly what you need.
Tell us the spec. We filter the existing library and deliver matching footage. No new collection required.
For data we don't already have, or for an ongoing stream, we run an assignment through our global contributor network.
An assignment is a structured spec. It defines what video to collect, who should collect it, and what makes a clip acceptable. Every parameter you set narrows or widens the contributor pool accordingly.
You describe the training objective in plain language, with examples, or by referencing an existing dataset. We use this to configure ML filtering and match contributors to the right tasks. You don't need to know every parameter upfront.
Geography, environment, perspective, activity, interaction type, device placement, speech. Each field filters the contributor pool. The system shows you how many contributors are eligible as you configure, so you can see the tradeoff between specificity and speed before committing.
Required elements, minimum duration, resolution floor, disallowed conditions. These become automated rejection rules enforced by the ML pipeline, not suggestions. You only see footage that passed every check.
Target is measured in accepted hours, not submitted hours. The difference matters: contributors typically submit 2.5 to 3 times the volume that ultimately reaches you, after pipeline filtering. Budget cap prevents overspend; the assignment pauses automatically if hit.
Tighter constraints mean higher signal and slower collection. Broader constraints mean faster throughput and more diversity. The right balance depends on your use case. We help you find it.
Narrow and specific. Smaller contributor pool. Higher per-clip relevance. Slower to fill but exactly on-distribution for a manipulation policy trained on kitchen tasks.
Balanced. Broad enough to fill quickly. Specific enough to be on-distribution. Most assignments land here, especially for scene understanding and action recognition tasks.
Every constraint you add narrows the pool. The system shows you the live impact as you configure, so you can decide whether a constraint is worth the slowdown before launching.
Our network is densest in North America, Southeast Asia, and Africa. Assignments requiring footage from lower-density regions take longer or require adjusted pricing.
Submissions don't go directly to you. They go through our ML pipeline first. Automated checks run in sequence, each one a gate. A clip that fails any check is rejected without reaching the next stage.
Format validation, resolution floor, audio track, file integrity, duration range, blur detection, exposure analysis, camera motion stability, occlusion check. Rejects corrupt, sub-resolution, or low-quality submissions on intake.
Verifies environment, perspective, and activity domain match the assignment spec. Uses object detection and scene models. A kitchen assignment rejects garage footage automatically.
CSAM detection and illegal-content classifiers run on every clip. Anything flagged is removed before delivery and never reaches your dataset. This stage is non-negotiable and runs regardless of assignment settings.
Clips that pass all stages are packaged with metadata and made available in your dataset. You see the acceptance rate, stage-by-stage breakdown, and constraint recommendations if the rate is below target.
Accepted footage arrives structured, documented, and ready to load. No cleaning pass required on your end.
Accepted clips in their original resolution, organized by assignment. MP4, H.264 or H.265. No transcoding unless requested.
Per-clip JSON with GPS, transcripts, detailed descriptions of activities, device type and more. Queryable and filterable.
Contributor agreements, license terms, and intended-use documentation for your legal and procurement teams. No ambiguity downstream.
Secure download, S3-compatible bucket transfer, or direct API. Datasets stream as footage is accepted, so you don't wait for the full assignment to complete.
Raw submissions that failed the pipeline. Contributor PII. Footage outside your spec. Anything without a complete rights chain. The pipeline exists so you don't have to deal with these.
For data we already have, the answer is days, not weeks. You send us the spec, we filter the library, and deliver a sample within two business days. Full delivery follows shortly after, depending on volume and your legal review timeline.
It depends on constraint specificity and volume. A moderately constrained 1000-hour assignment typically delivers in 7 to 10 days. Highly specific assignments (narrow geography, niche activity, head-mounted) can take 2 to 3 weeks at standard pricing, or faster with priority pricing. We give you a fulfillment estimate before you commit.
Yes. If you're building on top of Ego4D, EPIC-Kitchens, or another established benchmark, you can describe the distribution and we'll either filter our existing library to match it, or configure a new assignment taxonomy. You can also upload reference clips and our pipeline uses them for similarity scoring.
Contributors sign a rights-clearance agreement before recording. Contributor PII is never included in delivered datasets. We hold it internally for payment and compliance purposes only.
Every contributor signs a contract before recording that grants Syntropi a commercial license to the footage, which we sublicense to you. The license covers training AI models, including commercial products. We provide a per-assignment rights package for your legal team. We can discuss specific indemnification terms on a call. Our full privacy practices are documented at https://syntropiai.com/privacy.html.
Yes. Several of our customers run continuous assignments to feed training pipelines on a recurring basis. These are structured as monthly dataset streams: a standing assignment with a target per month, reviewed and adjusted quarterly. Typically better value than one-off assignments and easier to budget for.
We'll review your use case and send representative sample data within two business days. Receiving sample data does not commit you to any future purchase — we're here to help you evaluate fit.