Architecture
Deterministic where it can be, reviewable where it can't
Data Foundations sits between your people and your data stack. Facts come from code that reads your estate the same way every time; judgement comes from AI agents whose output is a proposal, not a change. Everything it produces is yours, in open formats, for the platforms you already run.
How it is built
Deterministic core
Qlik and MicroStrategy parsers, column profilers, the lineage builder and the scoring engines are ordinary code. The same input gives the same answer, and every figure traces to a line of script or a profile run.
Reviewable agents
AI agents (Claude by default) design data models, propose quality rules, convert code, draft conversion plans and read documents. Their output is stored as a proposal with its reasoning, validated for completeness, and applied only when a person accepts it.
Queue workers
Long jobs — a 5,000-dashboard ingest, a physical model, a document read — run on a queue in resumable chunks with progress reported as they go, so a browser tab is never the bottleneck.
Governed metadata store
Projects, runs, lineage, models and results live in PostgreSQL with row-level security on every exposed table. Reads of large estates are paged and writes are batched.
Outputs in open formats
SQL, DDL, pipeline code, Amazon Quick definitions, Excel workbooks and Word reports. Your team owns them; nothing depends on a proprietary runtime.
Runs and versions
Every run, version and approval is recorded with who did it and when, which is what turns a delivery into evidence.
Security and data handling
Hosting and data residency depend on the engagement — ask us about your requirements.
Roles
Owner, admin, member and viewer roles per project, plus a client role limited to answering assessments and uploading inputs.
Least access
Database connectors are read-only. Uploads go to project-scoped storage with signed links, and every project route checks membership before it runs.
Audit log
Decisions and changes — assessment moderation, contract breaches and more — are logged per project for owners and admins to review.
What the models see
Agents are sent the metadata, profiles, scripts or document text needed for the task — not whole tables. A project can use its own model endpoint, such as Azure OpenAI, instead.