Store every shape
Tables, vectors, graphs and files (images, video, documents) through one interface. You choose where each dataset lives: a database for transactions, or files and object storage for scale.
Governed data for people and AI agents.
One place for tables, vectors, graphs and files, with roles, an audit trail, a real erase, and strict sandboxes where AI can analyse data without being trusted with the keys. This page shows the design we are building toward.
What it will do
Govern it
A catalog of what exists, roles (Owner, Contributor, Reader) and service accounts, quotas, an audit trail, retention, and an erase that can be proven.
Compute safely
SQL, jobs and pipelines, and isolated sandboxes where an AI agent can run code to make charts and derived datasets. The sandbox never holds a storage key.
Run anywhere
On a laptop with SQLite and local files, or on a server with Postgres and S3-compatible storage. The same interface either way.
The design, in three blocks
Governance: the only door
APICatalogsIAMAccess policiesQuotaLifecycleAudit
Compute
Query engineJobsPipelinesSandboxes
Reaches storage only through a short-lived, limited handle that Governance creates.
Storage: one client, two families
Database familyPostgres. Real transactions across tables, for people who need them.
File and object familyS3-compatible object storage, and files on disk (including SQLite). Atomic commits per dataset, built for size and portability.
Every backend offers the same four things: relations, vectors, graphs and blobs.
Rules we are building around
- Everything goes through Governance. Nothing reads storage on the side.
- Code written by an AI runs in a sandbox with no network by default, strict time and memory limits, and no storage credentials. It reads and writes through checked, logged calls.
- What a sandbox produces (a chart, a derived dataset) inherits the strictest classification of what it read, and its lineage is recorded.
- Vector search obeys the same permissions as the data behind it, so an embedding never leaks a row you could not read.
- Erase is a durable task that retries until it is finished and proves the data is gone. Audit keeps the proof, not the person's content.
- Data never calls the products that use it. It announces changes, and others subscribe.
Where we are
Exists in codeA storage engine library (Postgres and SQLite databases, a file store, vectors, memory, documents) and a first server, with tests. Neither is deployed.
Being decidedThe storage contract, the file formats to support, and how closely to follow open catalog standards.
NextShort experiments to prove the design (concurrent writers, a real erase, vector search on a small machine), then the first customer: Loamz Agents.
Loamz Data is part of Loamz, alongside Loamz Chat.