Training data for physical AI and creative AI
Video training data you can defend.
We source, annotate and evaluate multimodal video, then ship every dataset with a record of where it came from, who agreed to its use, and how good the labels are.
What we do
Five pillars, and what each one is for.
Start with the pillar that matches your problem. Each has a named product and a stated outcome, and you can add the others later.
- PILLAR 01
Rights, consent, & provenance
VeriPhi PassportUse it to
- Show a buyer, auditor or regulator where each clip came from and what you may do with it.
- Document your training data for the EU AI Act before a high-risk model ships.
- Catch altered or re-encoded files before they reach training.
Outcome
Defensible training data, ready for EU AI Act compliance.
- PILLAR 02
Data & annotation
VeriPhi DatasetsUse it to
- Buy footage that already exists, or commission capture to your specification.
- Train spatial, manipulation and world models on video aligned with IMU and depth.
- Search a dataset by action, object or scene instead of watching hours of video.
Outcome
Targeted training data delivered fast, with precise multi-modal alignment.
- PILLAR 03
Workflow auditor & advisor
VeriPhi JudgeUse it to
- Score and rank model outputs without a person reading every one.
- Check outputs against your own business rules before they reach customers.
- Find duplicates, repeated work and labelling mistakes in your pipeline.
Outcome
Optimized, compliant workflows with less rework and lower error rates.
- PILLAR 04
Benchmarks & standards
VeriPhi BenchmarksUse it to
- Compare a dataset or a model with a baseline you can point to.
- Re-run the same safety checks on every release.
- Agree on one ontology across teams so labels mean the same thing.
Outcome
Confident model evaluation, safety checks and continuous deployment.
- PILLAR 05
Tailored SLMs & Custom Engine
VeriPhi EnginesUse it to
- Run a small model on a device where a large one is too slow or too costly.
- Use a model that enforces provenance and policy rules inside your pipeline.
- Get a model built for one task, such as spatial reasoning or robot action.
Outcome
Task-optimized, low-latency models tailored to specific enterprise needs.
PILLAR 01
Rights, consent, & provenance
Traces every asset in a dataset back to its hardware and licensing origin.
Use it to
- Show a buyer, auditor or regulator where each clip came from and what you may do with it.
- Document your training data for the EU AI Act before a high-risk model ships.
- Catch altered or re-encoded files before they reach training.
Defensible training data, ready for EU AI Act compliance.
Capabilities
- Rights-cleared licensing chain, with consent records
- Hardware-rooted C2PA signatures
- Device fingerprinting & tamper detection
- EU AI Act data documentation
- Sensor-to-cloud audit trail
Traces every asset in a dataset to its hardware and licensing origin.
EU AI Act, Article 10. Regulation (EU) 2026/1744, the Digital Omnibus on AI, took effect on 27 July 2026. For the standalone high-risk systems in Annex III it moved the start of the requirements from 2 August 2026 to 2 December 2027. Article 10 itself did not change. It asks for documentation of where training data came from and how it was collected, labelled and checked, and that record is much easier to build while the data is being made than after a model has shipped.
PILLAR 02
Data & annotation
Rights-cleared, multi-modal data built for spatial and enterprise AI.
Use it to
- Buy footage that already exists, or commission capture to your specification.
- Train spatial, manipulation and world models on video aligned with IMU and depth.
- Search a dataset by action, object or scene instead of watching hours of video.
Targeted training data delivered fast, with precise multi-modal alignment.
Capabilities
- Off-the-shelf and custom data collection
- Targeted sourcing from 15+ vetted partners
- Egocentric and spatial-intelligence coverage
- Multi-sensor alignment (video, IMU, depth)
- Vision metadata at shot, scene, segment and full-video level, with indexing
- Action, object and physical-design labels
- Scene graph generation
Rights-cleared, multi-modal data built for spatial and enterprise AI.
One taxonomy, manipulation-v2.3, applied at five levels of detail: video, segment, frame, object, pixel. Egocentric, cinematic, sensor-synced and multi-camera.
PILLAR 03
Workflow auditor & advisor
Evaluates, aligns and optimizes model outputs and business rules.
Use it to
- Score and rank model outputs without a person reading every one.
- Check outputs against your own business rules before they reach customers.
- Find duplicates, repeated work and labelling mistakes in your pipeline.
Optimized, compliant workflows with less rework and lower error rates.
Capabilities
- Automated RLAIF & preference scoring
- Enterprise rules-engine optimization
- Issue detection, deduplication & fixes
- Workflow streamlining
- Expert-in-the-loop rubrics
Evaluates, aligns and optimizes model outputs and business rules.
PILLAR 04
Benchmarks & standards
Measures data, models and agentic systems against trusted standards.
Use it to
- Compare a dataset or a model with a baseline you can point to.
- Re-run the same safety checks on every release.
- Agree on one ontology across teams so labels mean the same thing.
Confident model evaluation, safety checks and continuous deployment.
Capabilities
- Real-time, time-synced benchmarking
- Internal & external safety baselines
- Best-practice ontology standards
- Deep data discovery & indexing
Measures data, models and agentic systems against trusted standards.
In practice: inter-annotator agreement (Cohen's and Fleiss' κ, target above 0.85) measured on every project, gold-standard clips per domain, and a QA evidence pack you can check yourself.
Who we work with
Teams training models on video.
Physical AI
Household egocentric video from head- and wrist-mounted cameras: cooking, cleaning, sorting, folding, vacuuming. Labelled for hand-object interaction and manipulation policy training.
Creative AI
Broadcast and cinematic footage with scene, object and action labels and natural-language captions, prepared for generative model training.
Spatial and multimodal
Multi-camera rigs synced to under half a second, video paired with IMU streams, LiDAR and 3D capture. Custom modality sets to your specification.
How it works
Start small. See a sample before you commit.
- STEP 1
Send your brief
Domain, activities, volume, modalities and timeline.
- STEP 2
Get a plan in 5 days
Availability, schema, quality bar and indicative pricing.
- STEP 3
Evaluate a sample
Your team tests it before any production commitment.
- STEP 4
Receive with proof
Every batch ships with QA scores, a provenance manifest and C2PA credentials.
FAQ
Questions teams ask before they start.
How are you different from large annotation vendors and annotation platforms?
Large vendors sell managed labelling at high volume, and platforms give your own team the tools to label. Our starting point is household and manipulation video, with a taxonomy (manipulation-v2.3) written for it, and we welcome other use cases. Every dataset also ships with a provenance manifest, so the paperwork is part of the delivery and not a separate project.
What is the difference between annotation and provenance, and do you do both?
Annotation describes what is in the footage: actions, objects, scenes, motions, contacts and lighting. Provenance records where the data came from: consent basis, origin, rights scope, version history, and a cryptographic hash that shows the files have not changed since they were sealed. They are separate jobs and separate pillars, and we deliver both on every engagement.
What is inter-annotator agreement, and what score should I ask for?
It measures how consistently different annotators label the same clips. A Cohen's kappa above 0.80 is generally read as strong agreement, and below 0.60 is not reliable enough for production. We aim for above 0.85 and put the results in the QA evidence package that comes with each delivery. When you compare vendors, ask for scores per project rather than a general promise of quality.
Do you cover sensor data and multi-camera datasets, or only single-camera egocentric video?
We cover both. That includes head- and wrist-mounted egocentric video, egocentric video paired with IMU accelerometer and gyroscope streams, multi-camera rigs of 2 to 6 cameras synchronised to within half a second, cinematic RGB video, and 3D capture using COLMAP structure-from-motion. If your project needs a modality that is not listed, tell us. We welcome new use cases.
What does Article 10 of the EU AI Act require for training data?
For high-risk systems, Article 10 requires documented data governance: where the data came from, how it was collected, how it was prepared and labelled, what assumptions were made, and what gaps or biases are known. Regulation (EU) 2026/1744, in force since 27 July 2026, moved the start date for the Annex III high-risk systems to 2 December 2027. The requirement itself is unchanged. Our provenance manifest records origin, rights scope, preparation steps and quality results, and each delivery can carry C2PA content credentials. That supports your compliance work. It is not legal advice.
Can you annotate footage we already have, or only footage you collect?
Either. We can annotate your existing footage against manipulation-v2.3 or a custom schema, and we can arrange new collection to your specification: activity types, modalities, participant diversity and geography. Provenance documentation and agreement scoring come with both. We can also run an annotation quality evaluation on its own, on data you have already labelled.
Still have a question? Tell us what you are building and we will answer it directly.
Scope a project
Tell us what you are building.
A couple of lines is plenty. A person on our team reads every message and replies within 48 hours.
