The Platform

Datum

The data platform underneath every dataset we deliver — schema, warehouse, cleaning, lineage, and export. You create projects, upload data, and follow each job from ingest to export; it asks you to decide only when QA flags something.

Core platform — runs behind every delivery
Datum dashboard: job status counts and projects
Why a Named Platform

A labeled file is not a dataset.

Annotation produces labels. A dataset you can actually train on needs a schema that doesn't drift, a cleaning pass that catches duplicates and outliers, a record of where every row came from, and an export in the format your pipeline expects.

Datum is that layer. It's the same infrastructure whether the annotations came from CoreLabel Labelix or Robotix — one schema authority, one cleaning pipeline, one lineage graph, one export engine, regardless of which studio produced the labels.

Labels alone
Raw file Annotated Handed over as-is
Through Datum
Ingest Schema-checked Cleaned Versioned Exported
Every dataset version is queryable, auditable, and re-exportable
Inside Datum

The screens behind every job

Captured from the running platform.

Datum upload screen explaining the four stages a job goes through
Upload Four stages — ingest, AI pre-label, QA check — then a decision only if QA flags something.
Datum agent job settings with the human-review confidence threshold
Agent job Tasks below the confidence threshold are escalated to human review.
Datum ontology editor adding an object term with affordances
Ontology Only objects become drawable labels; materials and affordances stay vocabulary.

What Datum Does

Five layers between a raw upload and a dataset your model can actually train on.

Schema authority

PostgreSQL 15 with Alembic as the sole source of schema truth for every entity in the platform — one migration history, not five drifting copies.

Data cleaning

Deduplication, schema validation, and outlier removal run through dbt and Great Expectations before a raw upload ever reaches an annotator.

Analytics warehouse

Throughput metrics, client reporting, and billing data land in ClickHouse — a columnar store built for the queries a growing dataset actually needs.

Lineage & provenance

Every dataset version is tracked through DataHub and Apache Atlas: what it came from, what changed, and when — the record you need when a client asks "where did this row come from?"

Export engine

One export path producing COCO, YOLO, Pascal VOC, CSV, or JSON for annotated task lists — the format your training pipeline already expects, not one you have to build around.

Built and tested, growing in production. Five export formats are live and dataset versioning is in use today. As with any infrastructure layer, the honest way to judge it is by what it exports for you — ask us for a sample export in your format during a consultation.
The Platform

Datum feeds both studios — and takes nothing from either.

Whether a label comes from a human working in CoreLabel Labelix on a text span, or from Robotix on a 3D bounding box, it lands in the same schema, gets cleaned by the same pipeline, and exports through the same engine. The studio is where the work happens; Datum is where it becomes a dataset.

Frequently asked questions

What does Datum do?

Datum turns labels into a dataset you can train on. It holds the schema, cleans the data, records lineage and exports it in your format, whichever studio (Labelix or Robotix) produced the labels.

Which export formats does Datum support?

COCO, YOLO, Pascal VOC, CSV and JSON. All five are live, and dataset versioning is in use today. Ask us for a sample export in your format during a consultation.

How is data cleaned before annotation?

Deduplication, schema validation and outlier removal run through dbt and Great Expectations before a raw upload reaches an annotator.

Can I trace where a row in my dataset came from?

Yes. Every dataset version is tracked through DataHub and Apache Atlas: what it came from, what changed and when.

When do I need to step in during a job?

Only when QA flags something. Each upload goes through ingest, AI pre-labeling and a QA check, and tasks below the confidence threshold are escalated to human review.

Still have a question? Ask the team.

Need your data warehoused, cleaned, or exported on your terms?

Talk to us about your training pipeline's format, and we'll tell you exactly how Datum gets you there.