1Ingest and redact
Your data arrives encrypted. PII redaction removes personal data, and access is limited to the people on your project.
Liveingest
Encryption, redaction and access controls run today.
How it works
Four steps from raw data to a checked dataset. The first two run today. The last two start with our first pilot projects.
The four steps
2Pre-label
Self-hosted open-weights models fill in a first pass of labels: Segment Anything for images and video frames, domain-tuned LLMs for text.
Livepre_label
The pre-labelling engine runs today.
3Expert review
Subject-matter experts check and correct each pre-label. Ambiguous items go to several reviewers who cannot see each other’s answers.
Nextsme_review
Starts with the first pilot projects.
4Agreement check and delivery
We measure agreement between reviewers with Fleiss’ kappa and Cohen’s kappa, send disputed items back for another round, then deliver the dataset.
Nextconsensus_qa
Planned for the first pilot projects.
Step 1 of 4
Use the arrow keys, or the buttons, to move between steps. The drawings are illustrative.
How much the models do
70–80%
of labels pre-filled by models before expert review
Design target, not a measured result. Source needed: this figure comes from our internal business plan and has not yet been measured on a client project.
- Pre-labelling models
- Self-hosted, open-weights
- Review
- Subject-matter experts, blind multi-review on ambiguous items
- Agreement measures
- Fleiss’ kappa, Cohen’s kappa
Security and governance
Controls on client data, and whether each one runs today.
| Control | What it does | Status |
|---|---|---|
| Encryption | Data is encrypted in transit and at rest. | Live |
| PII redaction | Personal data is removed before anyone reviews an item. | Live |
| Access controls | Access is set per project and per role. | Live |
| Audit trails | A record of who viewed or changed each item. | Next |
What runs today and what comes next
What runs today
- Pre-labelling engineSelf-hosted open-weights models, for example Segment Anything and domain-tuned LLMs, produce a first pass of labels.
- EncryptionData is encrypted in transit and at rest.
- PII redactionPersonal data is removed before anyone reviews an item.
- Access controlsAccess to data is set per project and per role.
What comes next
- Expert reviewer networkWe are recruiting subject-matter experts now.
- Pilot projectsTaking requests now. Pilots start once the first reviewers are in place.
- Agreement checksBlind review by several experts on ambiguous items, tracked with Fleiss’ kappa and Cohen’s kappa.
- Audit trailsA record of who viewed or changed each item.
- Paid plansManaged annotation, dedicated expert pods and evaluation sets.
Updated .
Tell us about your data
Pilots start once our first expert reviewers are in place. Send a request now and we will reply by email.