New: prepared statements, UPDATE and DELETE, and key lookups that seek — 0.143 ms through the gateway.Read the update

Use case · Research & AI

Research & AI

Research groups and AI teams need a store they can reason about: a specified plan language, deterministic encodings, and guarantees with proofs rather than promises. The AI engines follow the same verifier when they ship.

Workload shape

WritesDatasets, embeddings later
ReadsAnalytics, reproducible queries
TenancyGroups and projects
NeedReproducibility, proofs
Where the design fits

How the data lands

Design fit, not a customer claim: which parts of UltraFabric carry which parts of this workload.

  • Every query is a typed plan with one canonical byte encoding and a fingerprint
  • The verifier's core theorem is mechanised in Lean 4 with no unproved steps
  • Vector, text and graph operators are specified with per-tenant index partitions
In SQL

What it looks like

The same plan, re-run at the same snapshot, returns bit-identical results; its fingerprint names it in the audit log.

CREATE TABLE samples (id bigint PRIMARY KEY, run text NOT NULL,
  at timestamp, label text, score numeric(8,4)) WITH (storage = columnar);

SELECT run, label, avg(score), count(*)
  FROM samples GROUP BY run, label ORDER BY run;
Status

Built today, coming next

Everything in the first column is in the repository and tested; the second column is the roadmap in order.

Built today

  • Canonical plan encoding; equal plans share a fingerprint
  • Lean 4 proofs for eight rewrite laws and the core isolation theorem
  • A ten-million-plan differential campaign with zero divergences

Next

  • Vector search behind the verifier, measured on ANN-Benchmarks
  • Text search and graph traversal
  • Retrieval with verified grounding

Tell us about yours.

Early access starts with a conversation about the workload you would move first.