For the data team

Weigh every number that matters. The same certified number, every time.

X-ray vision for your data. The same certified number, every time.

The model finds. The engine weighs. A person on your team certifies. Same words, same number; no model picks it.

Book a pilot Request demo access Watch the pitch 1:20

Terazu is the public name. The product you will see still carries its working name, Ontos, until the next release.

  • Runs on your servers, read-only against your data
  • Only your people certify. Terazu never does
  • Outside AI tools see structure, never your rows
What were net sales for FY2026 P09?
  1. Read as net sales · FY2026 P09 · checked against the contract0.3s
  2. Resolved to Net sales · Daily sales · certified0.1s
  3. Computing through the agent · core.sales_daily · 11,224 rows1.1s
Resolved to Net sales · Daily sales · certified · pin 90aa7e94
4,564,960.60 SARcomputed now, through your agent
Certified by Priya Sharma, Data Steward · 15 Oct 2026
read as · FY2026 P09, 23 Aug to 26 Sep 2026source · core.sales_dailySQL ▾Check this answer
Three questions from the demo, as recorded in October, asked from the data team’s seat.

Reads your databases, Informatica Cloud and Airflow live, and BI models and SSIS packages from files you upload. Certified route measured on PostgreSQL so far.

  • PostgreSQL
  • SQL Server
  • Oracle
  • BigQuery
  • Power BI
  • IBM Cognos
  • Informatica
  • Airflow

What Terazu does

It reads what your data passes through. Your people decide what’s true.

What your company knows about its data is spread across systems, code and heads. A model finds it, the engine checks it, and your people sign what holds.

Read how it works

Where every number comes from

Each number traced hop by hop to where its data is born. Every hop names an owner; what it can’t reach is a blind spot.

Whether it’s right

It measures each number everywhere else your company computes it, and lists where they differ and by how much, certified and most-used first.

What’s buried in your code

Joins in your procedures, query history and BI models are evidence. A model drafts links from them; the engine tests each against the data.

Facts your people sign

Nothing is a fact until its owner signs. The certification keeps its words, so those words get that number again, with no model.

Told when something changes

Terazu re-reads your systems on your schedule. A renamed column under a certified number pauses it and sends it to its owner.

Used by people and agents alike

Someone who can only ask gets a certified number, an offer of one, or a refusal. Your data team gets working answers, marked uncertified.

The short version

Terazu in 80 seconds.

Why one question gets three numbers, how every answer carries its receipt, how AI agents ask Terazu over MCP, and where Terazu goes next.

Filmed in October, under the working name, Ontos. Since then, AI tools over MCP get a link to the receipt, not the figure, and a view-only member gets only certified numbers.

Watch the full walkthrough, 3:45

The problem

Three people, three numbers. Then the AI makes four.

Ask three people for last month’s sales and get three defensible numbers. An AI makes a fourth. Terazu never picks one: the person who owns the number certifies it.

The monthly report 3,807,294.42 Calendar September, net
The finance pack 4,564,960.60 Fiscal period 9, net
The trading dashboard 4,653,345.11 Fiscal period 9, gross: markdowns left in
The AI chatbot “About 4.2M” Said with confidence. No definition named

“What were last month’s sales?”

Net sales · FY2026 P09 certified
4,564,960.60 SAR
  • Meanssales after markdowns and before VAT, in SAR
  • OwnerPriya Sharma, Data Steward
  • Bound tocore.sales_daily, through the fiscal calendar
  • If it changesit goes back to draft until someone certifies it again

The first three are real: one simulated retailer’s warehouse, in SAR.

How it works

From a messy estate to an answer you can defend.

Scroll through what Terazu does, one step at a time. This is an illustration of the flow; the findings and answers further down come from the live demo.

The eight steps in detail
  1. 01

    Point it at your estate

    Terazu starts from what you already run: databases, a warehouse, BI models, Informatica Cloud and Airflow. Nothing to migrate, and no model to write first.

  2. 02

    It maps what is actually there

    A read-only agent in your network reads schemas, procedures, pipelines and, where you allow it, query history. What it can’t reach is a blind spot.

  3. 03

    It drafts what things mean

    A model, with read-only SQL and every statement recorded, drafts what each table and number means. The engine tests each relationship against the data.

  4. 04

    Your team certifies it

    A person on the data team reviews the evidence and signs. The certification is pinned and keeps its words, so it can’t change unseen.

  5. 05

    Ask in plain language

    In the certified words, the number comes back with no model, computed at the source. A question it can’t pin down is refused.

  6. 06

    It keeps up as things change

    Terazu re-reads your systems on your schedule. When something under a certified definition changes, it pauses and goes back to its owner.

What it found in the demo estate

Read the estate. Find what’s wrong with it.

Pointed at a simulated retailer in October, Terazu read five systems through one agent and surfaced what a good data team would. Each finding is in the demo.

systems mapped
5
assets mapped
146
links traced
184
ontos · estate reading · live demo6 findings
  1. differs

    Net sales for period 9: two places agree to the halala; core.sales_by_channel is 0.04 SAR higher, loaded from the same table by its own procedure.

    check · core.sales_daily ↔ core.sales_by_channel · measured through the agent See it in the demo: Checks › Check results
  2. where it’s born

    Net sales traced five hops back to the till lines and voids in the Oracle point of sale, landed by an Airflow DAG.

    thread · core.sales_daily ← POS.TILL_LINES, POS.TILL_VOIDS · 5 hops See it in the demo: Estate › Threads
  3. missing loader

    Five tables under net sales have no loader Terazu can see, among them the exchange rates every sale is converted at.

    no captured loader · core.fx_rate, core.product_category + 3 more See it in the demo: Estate › Threads · Where it can break
  4. covers apart

    Over all of history, the warehouse’s net sales and the Cognos model’s are 0.40% apart; for period 9 they agree exactly.

    measured apart · core.sales_daily ↔ Cognos 'Net Sales' (MeridianFinance) · every row each keeps See it in the demo: Estate › Threads · Where it can break
  5. blind spot

    The Cognos finance model is known only from its exported files. Nothing about it has been checked against the live system.

    imported manifest · Meridian Finance (Cognos) See it in the demo: Estate › Atlas · What Ontos can’t see yet
  6. asked, not certified

    Two numbers people asked for are worked out but not certified: refunds on returns and the cancellation rate. Each waits for one click to certify.

    Needs you · numbers asked for, not certified · 2 See it in the demo: Home › Needs you

Certified, not certified, or not yet

Every answer says what it stands on.

A certified number cites who signed it. An uncertified one says so, for your data team. Someone who can only ask never gets a model’s number.

When Terazu refuses, and why
Someone who can only ask, last test 0 wrong
  1. The words a person certified18 of 18that number, with no model
  2. Other words for it18 of 18accepted offers right
  3. Something close, but not the same72 of 72refused
  • Read, then checked. A model reads the question; the engine checks it against what’s certified. The certified words skip the model.
  • Computed at the source. The number runs live through the agent in your network. A language model can narrate it; it never computes it.
  • Refuses in the open. A question it can’t pin down is refused, naming who can fix it. Asked again, it moves up the list.
  • Checked where it’s made. Check this answer measures the number wherever else your estate computes it. Agreement corroborates; it never proves.

The record

What the last test showed, and what it didn’t.

A view-only member asked 162 times, on a test warehouse Terazu had never seen. Scored by script and a blind judge, 7 November 2026.

wrong numbers
0
got a number
36
refused
126
ontos · last test · 7 November 20263 limits
  1. what it showed

    0 wrong numbers in 162 asks. 36 got a number. The other 126 were refused.

    36 in the test’s own words · 36 reworded · 72 near misses · 18 offers accepted
  2. one warehouse

    The one before served one wrong number in 72 near-miss runs; the build tested here closes that route.

  3. not a person

    An agent played your data team and could certify 3 of 6 questions. A measure or a link can’t yet express the rest.

  4. offers need judgement

    A script that accepted every offer got another question’s number in 42 of 48 near-miss runs.

Who it’s for

Understand the business, however new you are.

For the analysts, engineers and stewards who get asked where a number came from, and the newcomer learning it all. It pays off where numbers disagree and must stay in-house.

Who Terazu is for, in detail

Does this sound like you?

Two or more? That’s the estate Terazu is built for.

01

A new analyst starts Monday

They need to know what “net sales” means here and who to ask. In Terazu it’s written down, certified, with its evidence.

02

Two numbers reach the board

Finance and the business bring different figures under the same name, and each has an owner who’s sure.

03

Old and new run side by side

A migration or a merger leaves two sets of systems, and nobody can say which reports still matter.

04

An assistant is about to go live

Copilot, a chatbot or an agent is about to answer questions on company data, and nobody can say which of its numbers are right.

Where it matters most

Banks and insurers, health plans, listed retailers and consumer brands, the public sector and utilities: anywhere a number has to hold up in front of a regulator, an auditor or a board.

Probably not for you yet if

  • you run on one platform, and its own assistant already answers what you need
  • your metrics live in one well-kept semantic layer, and nothing disagrees
  • your data lives only in Snowflake or Databricks: both connectors are coming*
  • you want ETL, dashboards or search over documents
  • nobody can own the definitions. Terazu would mostly refuse.

* Not in the product yet; see the roadmap.

Works with what you run

Keep your stack. Add the part it’s missing.

Your warehouse, BI models and pipelines stay where they are. Terazu reads them as evidence and adds definitions, certified by the people who own them.

Databases BI models Pipelines
Terazucertified definitions
People AI assistants

Your catalog

Keeps listing what exists. Terazu adds what it means, certified by the people who own it, and names what it can’t see.

Your BI models

Stay the home of their definitions. Terazu reads Power BI, SSAS and Cognos models and binds each definition to the tables underneath.

Your pipelines

Keep running. Terazu reads Informatica Cloud and Airflow live, plus your stored procedures and SSIS packages you upload, to see what feeds what.

Your AI assistants

Ask and build over MCP from the same facts, and get the same refusal as everyone else. They see structure and counts, never your rows or figures.

Where your people already work

Work from Claude Code, GitHub Copilot or Cursor.

Point Claude Code, Copilot or Cursor at Terazu over MCP. It reads how a question was understood and which definition answers it; the figure opens for the person.

An owner or admin makes its key. It can ask and run checks, never certify, and every call is on the record.

claude — ~/meridianCopilot Chat · agent modeCursor · agent

What were net sales in FY2026 P09, and is that the certified number?

ontos · ontology.answer MCP

Read as Net sales · FY2026 P09 · certified · pin 90aa7e94

Receipt ↗ the figure opens for you, not the model

Ontos read this as certified net sales for FY2026 P09 (pin 90aa7e94). The figure stays inside Ontos: open the receipt to see it, with its SQL and who signed it.

Security by construction

It runs where your data lives.

The control plane and the agent run on your own servers. Choose the model: on your servers, or a cloud provider under your own key.

Security and deployment in detail
Your network
Postgres SQL Server Oracle
Terazu agentbeside your databases, on a read-only account
Terazu control planedefinitions, evidence, receipts
Your modelon-prem, or a cloud provider under your key

Self-hosted, signed

Both parts run on your Linux host from signed images, each release with its SBOM and scan, under an offline license. It can run air-gapped.

Agent beside your data

It needs only read-only access to your sources, dials out to the control plane, and never opens an inbound port.

AI is read-only

The model can read metadata and propose. It can’t write to your databases, certify, or compute the number it reports.

Every decision audited

Who proposed, certified and confirmed, on which evidence. The trail is tamper-evident and streams to your SIEM.

Access control

You decide what the AI sees.

Every person has a role, and each certified definition can name which roles get answers. The assistant acts as the person asking, so it never sees more.

  • Sign in your way. Single sign-on through your identity provider, with its MFA. Your directory makes and ends accounts, and its groups map to roles.
  • Keep your data away from the model. Set a source or the workspace to metadata only, or run local models only, and nothing leaves.
  • Sensitive data is classified. NDMO’s four levels, flagged as data is read inside your network; a Restricted column’s values never reach a non-local model.
  • Outside AI tools see shape, not content. Claude Code and other MCP clients get structure and counts, never a row or a business figure.
  • Only people certify. Service accounts and API keys are refused, and the assistant can prepare the case but never sign it.
  • Everything is recorded. Every tool the assistant calls, and every refusal, lands in the audit trail.
Who can do what
Settings / AssistantFinance workspace

Always on

Certified answerseach with its receipt
Definitions, SQL and lineagemetadata, no rows
A small sample on request5 rows by default, for people who may see rows, never from a metadata-only source

Your admins decide

Drafting definitionsproposals wait for a person
Table previews10 rows by default, at most 50
Profilingstatistics from a sample
Probesone read-only query, at most 200 rows, recorded
Checksmeasure a number where else it’s computed, at most 8 reads a number (a share or ratio checks its parts too), recorded

Metadata only: no rows or stored values reach a model

The whole workspaceevery source at once
HR systemSQL Server · the assistant sees names and structure; answers go to the person
WarehousePostgres · data allowed, within the tools above

Never, whatever the settings

Certify a definition
Write to your databases
Turn on query history
Change its own settings

Whatever is on, the assistant acts as the person asking, and never sees rows they couldn’t.

Start a pilot

Start with the numbers you already argue about.

See it on your warehouse. Your people sign only what’s needed.

  • Ask us for the install kit and a license key.
  • Run it on a Linux host with Docker, with one read-only login.
  • Certify one number, then ask for it as a viewer.

Nobody outside the team has timed this yet. We’re looking for a first pilot company.

How a pilot runs
Book a pilot Request demo access

Or write to the Terazu team.