Curatrix fits between the tools you already use.
Bring in footage from where it already is. Curatrix labels it, checks it and pixelates faces. Your team downloads the finished data into the tools it already uses. Curatrix is a hosted service in the EU.
Picture of how Curatrix connects. Left, where your footage is: your computer, your cloud storage and your own scripts. Footage is brought in to Curatrix in the middle, which labels it, reviews it, pixelates faces and releases it, and where your team signs in with its company login. Right, where your team uses the result: the release is downloaded into your training pipeline and your ML tools, and your own systems get notified.
The data comes out of the annotation platform, where it is labelled and reviewed.
Works with what you already use
The products below are examples of what each standard reaches. They show compatibility, not partnerships.
Storage
Where your recordings live today. Bring them in from your computer, from a script or from your cloud storage.
- Upload in the browser
- Python script
Cloud storage that uses the S3 standard, for example:
- Amazon S3
- Cloudflare R2
- Google Cloud Storage
- MinIO
- Wasabi
S3 is the common standard for cloud file storage. We set up the copy with you and test the connection before the first import.
Sign-in
Your people use their existing company login. No new passwords to hand out.
- Email and password
Any company login that uses the OpenID Connect standard, for example:
- Microsoft Entra ID
- Okta
- Keycloak
OpenID Connect is the standard behind most company logins. Only people you have invited get in.
Training tools and formats
What your ML team loads the data into. Pick the format when you download.
- LeRobot
- Hugging Face Datasets
- COCO
- YOLO
- CVAT XML
- KITTI
- RLDS
Formats are prepared with every release where they fit the data. LeRobot and RLDS are for robot data.
Automation
Your own systems get notified when data is ready and can download it without anyone clicking.
- Notifications to your systems
- Downloads by script
- Python toolkit
Not connected today
Shared drives have no connection to Curatrix yet. If your files live there, we bring them in with you by upload.
- Google Drive
- Dropbox
Every connection at a glance
Open a row to see what it means for your team. The precise detail for your technical team is one click further.
Data in
Drag files into the browser. A lost connection does not mean starting again.
When to use
Footage sits on a laptop or a drive.For your technical team
Resumable multipart upload: large files go up in parts straight to object storage. After a reload the browser asks which parts are missing and sends only those.
Your own footage, labelled by our team and released only to you.
When to use
You want your own recordings labelled.For your technical team
An admin of your organisation uploads footage into a project commissioned to you. Our team labels it, and it is released only to you. It never enters a catalogue, a sample or another customer's licence.
A few lines of code send files without anyone clicking.
When to use
Footage comes out of a machine or a pipeline.For your technical team
The Python SDK declares a file, sends it straight to storage and completes the upload. A file of up to 5 GB goes in one request.
We copy files from your cloud storage, together with your team.
When to use
Footage already sits in cloud storage.For your technical team
S3-compatible bucket import: together with our team, you can import files from a bucket reached over HTTPS. The files are copied into Curatrix storage. The access key is stored encrypted and never shown again.
Identity and access
Your people sign in with the company login they already have.
When to use
Your company has one central login.For your technical team
Single sign-on with OpenID Connect: an admin connects your provider, matched by your email domain. Sign-in uses the authorisation code flow with PKCE, and the ID token is checked against your provider's keys.
Only people you have added get in. Nobody gets an account by accident.
When to use
Always, it is how sign-in works.For your technical team
Single sign-on admits people who are already members. No account is created on first sign-in, and nothing is provisioned automatically.
Admins invite colleagues by email and decide who may release data.
When to use
A colleague joins or leaves.For your technical team
Admins invite people by email, change their role and remove them. Members do the work. Admins sign off privacy reviews, release datasets and manage keys.
Scripts get their own key, which an admin can switch off at any time.
When to use
A script uploads or downloads.For your technical team
Service keys: scripts use a key of your organisation. An admin creates it, it is shown once and stored as a hash, and it can be revoked at any time.
Automation
Your developers fetch releases and files from code instead of a screen.
When to use
You connect your own tools.For your technical team
REST API with OpenAPI: releases, assets, acceptance and webhooks sit under /api/v1. The OpenAPI 3.1 document is served next to it.
Each key does only the job it was made for.
When to use
You hand a key to one system.For your technical team
Scoped keys: a key can be limited to reading releases or assets, recording acceptance, managing webhooks or requesting samples.
Your systems get a message as soon as new data is ready.
When to use
Nobody should have to check for new data.For your technical team
Signed webhooks with retries: five events, a release published, licensed to you or accepted, a job done and a sample ready. Each delivery is signed with HMAC-SHA256. A failed delivery is retried with the same event id, eight attempts in total over about 18 hours.
If a request is sent twice, you still get one result, not two copies.
When to use
Networks drop and scripts retry.For your technical team
Idempotency keys: calls that create an upload, a job or a dataset accept one. A repeated call returns the first result instead of a second copy.
Data out
Every download checks your licence first and comes straight from storage.
When to use
Every download.For your technical team
Short-lived links: every download is checked against your licence first. The link is valid for five minutes, and the file comes straight from storage.
Pick the format your tools already read.
When to use
Your team trains with a specific tool.For your technical team
You choose the format when you download: the main archive, or a LeRobot v3.0 export or a COCO, YOLO, CVAT XML, KITTI, RLDS or Hugging Face label pack where the release offers one.
Your pipeline can prove that every file arrived complete and unchanged.
When to use
Before training on a release.For your technical team
Checksums and a signed manifest: the manifest lists a SHA-256 checksum for every file and is signed with Ed25519. Your pipeline can verify the archive before training on it.
Tell us a release is accepted, or exactly what is wrong with it.
When to use
After checking a release.For your technical team
Your pipeline confirms ingestion and accepts a release through the API, or rejects it with a structured issue that the supplier resolves.
What your security team gets
The controls behind these connections are described in detail on the trust page.
- Organisation isolation
- Every record belongs to one organisation. A request for another organisation's data is answered as not found.
- Append-only audit log
- Role changes, privacy sign-offs, releases and deletion requests are logged per organisation and cannot be changed.
- Face review before release
- Faces are detected, corrected by a person and pixelated. A release is refused until every video is cleared.
- Deletion
- When an admin deletes a recording, access ends at once and a worker purges every stored version.
Writing the integration?
The developer page explains the API, the Python SDK, webhooks and how access keys work.
Walk through your architecture with us.
Bring your storage, identity and pipeline setup. We show where each connection sits and what your team sets up on its side.
