Custom data collection for AI training, with consent

Images, video, speech and text collected to your spec, with written consent, balanced quotas and metadata on every item.

Start with a pilot batch See what we collect

example spec · shelf photos · 3 lighting conditions

Signed consent on fileevery item Metadata: device, lighting, locationcomplete Blur, duplicates, wrong framing2 rejected

What we collect

New data, made to your specification, with consent and the metadata your model needs.

Image collection

Photos of objects, scenes, products, documents or people, across the devices, angles and conditions you specify.

Used for: Detection, recognition, retail, OCR

Video collection

Clips of actions, gestures, driving or in-store scenes, filmed to a shot list.

Used for: Action recognition, tracking, robotics

Speech recording

Scripted prompts or natural conversations, by speakers you select by language, age band and region.

Used for: ASR, wake words, TTS

Text and conversation data

Written prompts, dialogues and domain text produced by people, not scraped.

Used for: Chatbots, LLM fine-tuning, NLP

Document samples

Receipts, forms and handwritten notes, created for the project or shared with documented permission.

Used for: Document AI, handwriting, KYC flows

Edge-case capture

The rare situations your model gets wrong: low light, glare, occlusion, unusual objects.

Used for: Fixing known model failures

Balanced sampling

Quotas across lighting, location, device, accent or other factors, tracked as items come in.

Used for: Fairer, more robust models

Collection plus labeling

Collected and annotated in one project, with one guideline and one QA report.

Used for: End-to-end datasets

What goes in a collection spec

A good spec decides most of the dataset's quality. We write it with you before anyone collects a single item.

Spec item	What we agree	Example
Volume and quotas	Total items and the split across conditions	Even split across three lighting conditions
Contributors	Who can take part, and the mix you need	Adults only; mix of regions and age bands
Capture rules	Devices, angles, distance, duration, background	Phone camera, product fills 60% of frame
Quality checks	What makes an item rejected	Blur, duplicates, wrong framing, clipped audio
Metadata	Fields recorded for every item	Device model, lighting, city, timestamp
Consent and rights	Consent form, usage rights, withdrawal process	Written consent per contributor, rights in contract
Delivery	File format, naming, folder structure	JPG + manifest CSV with consent ID per file

Consent and ethics

How data is collected matters as much as what is collected. These rules apply to every project.

Written consent. Every contributor signs a consent form that explains the purpose in plain language.
Right to withdraw. Contributors can withdraw, and their items are removed from future deliveries.
Adults only. We do not collect data from anyone under 18.
No scraping of personal data. Collected data is created for your project, not copied from the internet.
Fair pay. Contributors are paid for their time, and the rate is agreed before they start.
Clear rights. Ownership and usage rights for the dataset are written into the contract.

How a project runs

Pilot first, scale only after you have seen real data.

Agree the spec
Volume, quotas, capture rules, metadata, consent and delivery format.
Recruit contributors
Directly or through vetted partners, matched to your quotas.
Pilot collection
A small first batch, checked item by item against the spec.
Refine
You review. We adjust capture rules and quotas.
Scale
Full collection in batches, each with a QA and quota report.

What each delivery includes

So you can trust every file, and trace it back if you ever need to.

The dataFiles in the agreed format, named and organised as specified.
A manifestOne row per item with metadata and a consent ID.
Quota reportProgress against every quota, so gaps are visible early.
Rejection logItems rejected, with the reason.
Consent recordsStored securely and available for your audit on request.

Data security

Collected data often includes faces and voices. It is handled as personal data from day one.

NDA and contract first. Signed before collection starts.
Need-to-know access. Only the project team can open collected files.
Secure transfer. Delivered through a secure link or directly into your storage.
Retention you control. Working copies deleted after delivery, or on the schedule you set.

Where teams use it

Most collection projects start because public data does not match the real world.

Computer vision

Products, shelves, scenes and objects in the conditions your users actually face.

Speech and voice

Accents, languages and noisy environments that public datasets miss.

Document AI

Receipts, forms and handwriting for extraction models.

Retail and e-commerce

Product photos across angles, packaging and lighting.

Driving and robotics

Edge cases and rare scenarios on a shot list.

LLM and chat

Human-written prompts and conversations for fine-tuning.

Questions teams ask

If yours isn't here, ask in the form below.

How is data collection priced?

Per item or per hour of recording, depending on how hard the items are to capture, the contributor profile and the metadata you need. We quote after agreeing the spec and confirm after the pilot.

How do you find contributors?

We recruit for each project, directly and through vetted partner teams, and confirm we can reach your quotas before quoting.

Can you collect in specific cities or languages?

Usually, yes. Tell us where and in which languages, and we confirm reach before the pilot rather than promising it upfront.

Who owns the collected data?

Ownership and usage rights are written into the contract before collection starts.

Can you label the data too?

Yes. We can collect and annotate in one project. See image & video, text & nlp, audio & speech, llm evaluation.

Do you sell ready-made datasets?

No. Every dataset is collected for your project, to your spec.

Start with a pilot batch

Tell us what data your model needs and where it falls short today. We'll reply with questions and a draft collection spec.

Prefer email? Write to hello@norbeon.com.

Run an annotation or data team? Register as a vendor.