Where it’s going, and what works today.

In the product

• A census of your estatePostgres, SQL Server, Oracle, BigQuery, Informatica Cloud and Airflow read live; Power BI, SSAS, SSIS and Cognos from files you upload. Blind spots are named.
• Informatica Cloud and Airflow, read liveTaskflows, mappings and DAGs, and what each run’s log shows it reading and loading, kept as lineage evidence. Nothing in either is changed.
• Threads and ownersFollow a report back to where its data is born: who owns each hop, how Terazu knows it, where it can break. People declare owners.
• Definitions drafted from evidenceWhat each table and number means, drafted from schemas, SQL, BI models, pipelines and, where you allow it, query history, with its grain and evidence.
• The analyst, for your data teamA model writes read-only SQL through the recorded probe, every statement on the record. The engine checks each relationship. Replies are working answers until certified.
• Disagreements measuredWhen two systems claim the same meaning, Terazu measures how far apart they are and puts it to their owners.
• Certification by named peopleOne signature certifies a definition, pinned to its exact version; a workspace can ask for a second. A change underneath sends it back to review.
• Needs you: certification made quickDrafts sorted by the decision they need, each with a recommendation and its evidence, settled in one click or in bulk. Sorting the pile
• From working answer to certifiedMake this certifiable turns a working answer into cards, tries them, and shows whether the served number agrees before a person certifies.
• Same words, same numberA certification keeps its reading, so the same words get the same number with no model. Reworded, a model only picks and the engine checks.
• Reports as a demand signalWhich reports read each table and how often they run, counted with the questions people ask, to decide what to certify next.
• Domains, when you want themProposed only from your BI models and report folders, each accepted by a person. Optional, and only in the Everything app mode.
• Every answer says what it stands onA certified number cites its definition, who signed it, the source and the SQL. An uncertified one says so, and reaches only your data team.
• The question, read and checkedThe engine checks the model’s reading against what’s certified. Someone who can only ask gets a certified number, an offer they accept, or a refusal.
• Check this answerMeasured wherever else your estate computes it: where they agree, where they part, and why. Agreement corroborates; it never proves. When two numbers disagree
• Check every numberEvery Measure checked for one period, certified and most-used first, in one ranked list of what differs. The list never picks a side.
• A ruling changes the definitionA decision made in the chat carries its changes. Approving it applies and certifies them, as that person, and answers cite the ruling.
• Numbers built from certified MeasuresA difference, share, ratio or change worked out from certified Measures, each part’s pin shown, labelled not certified until Certify this formula.
• Any period, and comparisons“Last fiscal year”, “Q2”, “this month against last”, from a calendar certified once. “Today” is the workspace’s day, in its time zone.
• The evidence behind every LinkEach Link Terazu holds keeps its evidence. A Link certified on a person’s word keeps their sentence and the engine’s verdict beside it.
• Roles, audiences and a governed assistantFive roles per workspace, invites, audiences per definition, and certification by people only. Every change the assistant proposes is a card a person approves.
• Metadata only, whole workspace or source by sourceNo rows or stored values from a metadata-only source reach any model, in the chat or over MCP. A certified number still reaches the person.
• Ask and build from Claude Code, GitHub Copilot and CursorWith their own key, they get structure and counts, never a row or a figure. Every call is recorded, and they can’t certify.
• Sign-in through your identity providerSingle sign-on with your provider’s MFA. Your directory makes and ends accounts over SCIM, its groups mapped to roles.
• Ready for a security reviewAn audit trail streamed to your SIEM, NDMO classification, screening for personal, health and credential data, and a model’s SQL reads only catalogued tables.
• Query history, only where you turn it onOff until a person turns it on, per connection. Statements and run counts: a person is never named; a load or copy job’s service account is.
• Runs on your serversSigned images with an SBOM, an offline license, read-only access, air-gapped if needed. Measured: a probe in 70 ms at 21,000 tables.