How it works

From the estate you run to numbers your team can defend.

Terazu reads what you already run and drafts what things mean from that evidence. Your data team certifies just enough to answer the questions that matter.

Book a pilot Watch the walkthrough

Watch it run

One retailer, end to end.

A narrated walkthrough of the hosted demo, under four minutes: a simulated retailer with an Oracle point of sale, a Postgres warehouse, Airflow and Informatica Cloud.

The walkthrough is being recorded

Until then, ask for an access code and try the demo yourself.

Request demo access

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

Wherever you’re starting from

No model to write first.

Most estates have some of these layers, often out of date. Terazu starts from what exists, and you pick the starting point from the census.

Semantic first

A mature BI model

Power BI or SSAS models already hold measures. Upload them, and Terazu traces each measure to the tables underneath and checks it against the sources.

Model first

A logical model

Bind its entities to the tables that implement them, compare with how they’re used, and propose the measures it’s missing.

Mart first

Data marts only

Build one contract across the marts, and show where they disagree.

Evidence first

Databases, and not much else

Start from schemas, keys, SQL, procedures, Informatica and Airflow runs, and data profiles, plus query history where you turn it on.

No warehouse

Operational systems only

Start from the operational databases, with lower confidence and more human review.

The lifecycle

Eight steps, from census to a contract that keeps up.

People on the data team make the calls: which questions come first, which definition stands, what gets certified. Everything else is evidence.

  1. 01
    Connect · Discover

    See what exists

    A read-only agent beside your databases maps systems and pipelines; uploaded BI models join the map. What it can’t reach is named, not hidden.

    In the product: Estate › Atlas, and Estate › Threads to follow a report back to its sources.

    Estate AtlasInventory and evidence only. Dark territory is named, not hidden.
    5Systems
    4Connections
    146live Assets
    184Asset edges
    100%traced to a source
    What the lines say
    reads102
    uses37
    writes25
    joins13
    derives from7
    How we know
    46 declared by the source 119 read from SQL, runs and logs 19 from imported manifests
    Dark territory: 1 named blind spot
    • Meridian Finance (Cognos) · known only from its exported files, never checked against the live system
  2. 02
    You decide

    Pick the questions first

    A person picks the questions that matter most, using the census. No domain needed: Terazu reads the whole estate within your read budget, most-used first.

    In the product: Home lists the questions people ask, from query history, and Coverage shows which answer. Domains are an optional lens a person accepts.

    Questions people already ask5 picked first, by a person
    • Net sales by fiscal period
    • Net sales by district
    • Ramadan against last Ramadan
    • Refunds on returns
    • Gross margin
    Focus · optionala Domain is a lens, never a boundary
    No focus · the whole estate Sales Stores Products Returns Online orders

    An illustration, with questions asked on the demo.

  3. 03
    Discover

    Harvest meaning where it already lives

    Highest-value evidence first: BI models that define measures, then the most-run reports and queries, then transformations and lineage, then schemas and profiles.

    In the product: the Atlas, query history in Connections (off until you turn it on), and uploaded BI models, SSIS packages and Cognos models.

    Evidence · highest value firstthe demo’s workspace
    1. 1BI modelMeridianFinance (Cognos) · its measures transcribedimported
    2. 2Queries people runquery log · 32 query patternscaptured
    3. 3Pipelines and transformationsInformatica mappings, Airflow DAGs, 35 proceduresread live
    4. 4Schemas49 tables, declared foreign keys, key domainscaptured
    5. 5Profilesrow counts, null rates, value overlap · 43 tablescaptured
    TracedNet Sales = total([SalesWeekly].[NET_AMT])legacy.FACT_SALES_WEEKLY.NET_AMT

    The demo’s capture state. The Cognos model’s measure is traced to the table underneath it.

  4. 04
    Govern

    Draft the contract

    A model drafts what each table and number means, every statement recorded. The engine checks every relationship. Until a person certifies, it’s a working answer.

    In the product: Ontology › Objects › Needs you, drafts sorted by the decision they need, and Ontology › Measures.

    draftNet salesMeasure · Daily sales
    formula
    SUM(net_sales), money in SAR
    grain
    one row per store, product and day
    served from
    core.sales_daily
    evidence
    Cognos measure ‘Net Sales’checked: 129,997,505.75 over 322,417 rows

    Not certified yet: an answer from it says so.

    The demo’s Net sales, in its draft state, checked before a person certified it.

  5. 05
    Govern

    Show where definitions conflict

    Where systems define a number differently, Terazu measures the gap and puts it on the record. It never picks a winner; a person does.

    In the product: Check every number and its Numbers that differ, Checks › Check results, Estate › Threads under “Where it can break”, and Check this answer.

    Check · net sales, FY2026 P09differs
    The certified numbercore.sales_daily · 11,224 rows4,564,960.60
    A summary tablecore.sales_by_channel · 225 rows4,564,960.64
    0.04SAR apart

    Two other places agree to the halala. No winner picked: the difference goes on the record, and the people who own the number decide.

    From the demo’s Check results, measured through the agent.

  6. 06
    People certify

    Certify what it means

    Each review carries evidence and a recommendation, so obvious ones are certified a group at a time. One person signs, and the content is pinned.

    In the product: Ontology › Objects › Needs you, and Ontology › Measures, where a draft is checked against its BI model first. From the demo:

    MeasureDefinitionState
    Net salesDaily salesSUM(net_sales)certified15 Oct 2026pin 90aa7e94
    Gross salesDaily salesSUM(gross_sales)certifiedpin a2bc1542
    Refunds on returnsSales return · asked 3×SUM(refund_amount)not certifiedkept from a question
    Cancellation rateOnline ordershare with status ‘CANCELLED’not certifiedkept from a question
    Gross marginaskedno definitionnot defineda stated “not yet”
  7. 07
    Serve

    Certify just enough, then serve it

    Certified words get their number with no model. Someone who can only ask gets that, an accepted offer, or a refusal naming who can certify.

    In the product: the chat and Ontology › Coverage. Your data team also gets working answers, marked uncertified. What every answer stands on

    A person · the chat
    What were net sales for FY2026 P09?
    Resolved to Net sales · Daily sales · certified · pin 90aa7e94
    4,564,960.60SAR, computed now, through your agent
    read as · net sales · FY2026 P09certified by · Priya Sharma (Data Steward), 15 Oct 2026SQL ▾
    How much gross margin did we make in FY2026 P09?
    Not yet. Gross margin isn’t a certified Measure.

    Nothing was estimated. Ask it as gross sales or net sales.

    needs “gross margin”
    An AI assistant · MCP 
    → ontology.answer  “What were net sales for FY2026 P09?”← read as    Net sales · FY2026 P09  resolved   certified · pin 90aa7e94  receipt    link ↗  figure     opens for the person, never the model → ontology.answer  “How much gross margin did we make in FY2026 P09?”← not yet    needs “gross margin”

    The demo’s receipt and refusal from October, simplified. Over MCP the figure stays with the person; certified readings aren’t served there yet.

  8. 08
    Ongoing

    Keep up as things change

    Terazu re-reads stale evidence on your schedule. A breaking change under a certified definition sends it back to review, and readings on it stop serving.

    In the product: re-certification, and Home › Needs you, where numbers asked for but not certified wait, most asked first.

    Breaking change · core.sales_daily.net_sales renamed
    Net salespin 90aa7e94certifiedneeds review

    Answers from it pause, and it goes back to review before anyone sees a new number.

    Needs you · numbers asked for, not certified
    refunds on returnsthe total of refund_amount on Sales return · kept from a questionasked 3×

    The change is an illustration; the list is the demo’s.

When nothing certified covers it

Questions pull context in, too.

When nothing certified covers a number, Terazu works it out for your data team from certified pieces, marked not certified. Someone who can only ask is refused.

Worked out once, kept for next time.Not certified until a person says so.
When Terazu refuses, and why
ontos · a number not certified yet · demo, October
  1. asked

    “How much did we refund on returns in FY2026 P09?” No certified Measure computes refunds.

    no certified Measure · the question names refunds and a period
  2. read

    A model reads it as the total of refund_amount on Sales return over period 9. The engine checks each piece is certified, calendar included.

    read as · total of refund_amount · Sales return · FY2026 P09
  3. answered

    25,307.45, computed through the agent and marked Not certified: “the total of refund_amount on Sales return, kept from this question”, with Certify beside it.

    answer · not certified · the same governed SQL a certified number runs
  4. kept

    The definition is kept as a draft Measure. Asked again, it answers the same way, and still says it isn’t certified.

    draft Measure · kept from a question · 15 Oct 2026
  5. listed

    It waits on Home’s Needs you, asked three times so far. Once a person certifies it, the data team gets it as certified from the next question on.

    Needs you · numbers asked for, not certified · Certify

Start a pilot

Start with the numbers you already argue about.

Steps one to seven, on your estate, for the numbers that matter most. Terazu runs on your servers, read-only against your data.

Book a pilot Request demo access

Or write to the Terazu team.