Data curation
Your data. Ready for AI.
We run our curation pipeline on your documents, scans, audio and video. On our own machines in Germany under a data processing agreement, or inside your environment.
What goes in
- PDFs, including scanned pages
- Office documents, HTML and e-mail exports
- CSV files, logs and database exports
- Audio and video recordings
What comes out
- Structured JSONL or Parquet in an agreed schema
- Personal data removed or masked according to your policy
- Language, quality flags and content hash on every record
- A data card and a before-and-after quality report
The pipeline
Eight steps from raw files to training data.
The same pipeline that builds our own corpora. Each step can be switched on or off for your project.
Extract
Text is pulled out of documents. Scanned pages go through optical character recognition.
Transcribe
Audio and video are converted and transcribed to text.
Clean
Encoding damage, markup, page furniture and boilerplate are removed, and text is normalised.
Scrub personal data
Names, e-mail addresses, phone numbers, bank details and identifiers are detected and removed or masked.
De-duplicate
Exact and near copies are found, so nothing is over-weighted in training.
Detect language
Independent detectors vote, and a label is kept only when they agree.
Flag quality
Every record is checked. Low-quality records are flagged or removed, as you decide.
Structure and report
Output in your schema, with a data card and a benchmark report.
What we do not promise: automated personal-data detection does not find everything. We report detected rates and known limits, and for sensitive data we agree a review step with you before delivery.
Formats and services
Delivered the way your pipeline expects.
Fine-tuning data
Instruction, chat, preference (DPO) and completion formats, matched to the framework you train with.
Pre-training data
Columnar Parquet, ready for Hugging Face Datasets, Spark, S3 or BigQuery.
Retrieval (RAG)
Segmentation, chunking and metadata tuned to your vector database and embedding model.
Evaluation sets
Held-out sets built to your rubric, delivered with a scoring script.
Recurring pipelines
Repeatable batch jobs with automatic quality reports, for teams with a steady flow of data.
White-label
Consultancies and data vendors can offer the pipeline to their own clients under their own name.
Data protection
Built for data that has to stay under control.
Agreement first
A data processing agreement under Art. 28 GDPR is signed before any data is touched. An NDA is standard.
Processed in Germany
On our own machines in Germany, or on yours. Your data is not uploaded to cloud storage.
Encrypted at rest
Data is held encrypted for the duration of the project only.
Deleted after delivery
Your source data is deleted when the project ends, and you receive written confirmation.
Two ways to run it
On our machines or on yours.
| Aspect | On our machines | In your environment |
|---|---|---|
| How it works | You send the data over an encrypted transfer. We process it in Germany and delete it after delivery. | We run the pipeline inside your environment, remotely on a machine you provide or on site. |
| Suited to | AI teams, software companies, and data that may leave your network under contract. | Banks, insurers, healthcare, pharma and public bodies whose data may not leave the building. |
| Your data | Held encrypted for the duration of the project only. | Never leaves your infrastructure. |
Who it is for
Organisations with archives AI cannot use yet.
Banks and insurers
Contracts, correspondence and case files turned into training and retrieval data without personal details.
Healthcare and pharma
Reports, study documents and recorded speech prepared for domain models under strict data protection.
Public sector and legal
Scanned records, rulings and administrative documents made searchable and usable.
Consultancies and data vendors
The pipeline under your name, for your own clients.
How a project runs
Start small. Then scale.
Scoping
A call about your data, formats, volumes and data-protection needs.
Pilot
We curate a sample of about 1,000 records and show the result with a quality report.
Quote
A fixed price based on what the pilot showed, not on a guess.
Full run
The whole dataset goes through the pipeline, with interim figures as it runs.
Delivery
Data, data card and report are delivered. Your source data is deleted and confirmed.
Questions about data curation.
Does our data have to leave our premises?
No. The pipeline can run inside your environment, on a machine you provide. If you prefer, you send the data to us under a data processing agreement and we process it on our own machines in Germany.
Do you sign a data processing agreement?
Yes. An agreement under Art. 28 GDPR is signed before we receive any data, together with an NDA.
Which file types can you process?
PDF including scans, Office documents, HTML, plain text, CSV and database exports, and audio and video files. Scans go through OCR and recordings through speech-to-text.
Which languages are supported?
Language detection covers a very wide range of languages, including low-resource ones, because the pipeline was built for a corpus of more than 130 languages. OCR and speech-to-text quality varies by language, which is one reason we start with a pilot.
How is a project priced?
Per project, by volume, formats and the way it runs. The pilot gives both sides real numbers, and the quote follows from it.
Send us a description of your data.
Formats, rough volume and what you want to use it for. We reply with a pilot proposal.