An open lakehouse you can walk away from.

Snapshots, schema evolution and point-in-time reads. A query against last Tuesday returns what was true last Tuesday, not what is true now.

Tables live in the open Apache Iceberg format on S3-compatible storage, readable by any engine. Schema evolves without a rewrite, and a point-in-time read returns what was true then — the property that makes leaving a copy, not a project.

Trino and Spark over the same tables, with pipelines that declare their lineage rather than having it inferred afterwards.

The engines you already know run over the same governed tables, and a pipeline declares its lineage as it runs — so where a figure came from is recorded while the work happens, not reconstructed for the audit.

Experiments, a model registry, a feature store with point-in-time correctness, and promotion gated on a model card and an approval. Every served model traces back to the snapshot it was trained on.

Features are point-in-time correct, so training never reads the future. A model reaches serving through a model card and an approval, and what is served always traces to the exact snapshot it trained on.

Retrieval over your own governed data, with output schemas, budgets, a sovereign mode that keeps inference in-region, and the same entitlement filtering as every other read.

The model reads only what the person asking is allowed to see, because retrieval goes through the same entitlement filter as every other read. Output schemas, budgets and in-region inference are properties of the run, not settings to remember.

Every table registered, every movement recorded as an edge, and sharing that is a grant rather than an export.

The catalog registers every table and records every movement as an edge, so a number on a dashboard traces back through every step. Sharing is an auditable grant — never a copied file nobody can recall.