Skip to content
XYPHORS
A — Data & AnalyticsService 03 · B3

One model ofthe truth.

Pipelines, warehouse modelling and reporting that reconcile. We start from the decisions a team needs to make, work backwards to the metrics that inform them, and only then to the tables underneath.

Bring a reporting problemAll services
X
LineageWhere a number came from, and what derived it.
Y
MeaningWhat it is supposed to represent, and who owns that.
B — Working orderDecision → table

Reporting projects fail at the top of the ladder, not the bottom.

The tables are rarely the hard part. What breaks a data project is starting there: modelling whatever the source system happens to emit, and discovering afterwards that nobody agreed what the numbers were supposed to mean.

So we work upward. Which decision, made by whom, how often. Then the metric that informs it, written down with its edge cases and an owner. Only then the model, and only then the tables.

Working order4 steps, upward
  1. 01
    Decisionwho chooses, how often
  2. 02
    Metricdefined, owned, written down
  3. 03
    Modelentities, grain, history
  4. 04
    Tablethe last thing designed

Read upward — the table is derived, not assumed

C — The pipelineSources → insight
Pipeline and reconciliationC — figure 01
PIPELINE01SourcesAS THEY ARE02IngestSCHEDULED03TransformTESTED04ModelGRAIN + HISTORY05InsightA DECISIONRECONCILEDONE ANSWER
Station with a defined outputSeries before reconciliationInset — one definition
Five stations, each with an output that can be checked. The inset is the point of the exercise: two series that disagreed, brought onto one line once the definitions were settled and owned.
  1. 01

    Ingestion

    sources, as they are

  2. 02

    Pipelines

    scheduled, observable

  3. 03

    Transformation

    tested, versioned

  4. 04

    Models

    entities and metrics

  5. 05

    Insight

    a decision someone makes

D — SymptomsWhy teams call us
  1. 01

    Two dashboards, two answers

    Almost always a definition problem rather than a query problem. The fix is a written metric definition with an owner, and then the model that enforces it.

  2. 02

    A simple question takes a week

    When every answer needs an analyst to join four exports by hand, the bottleneck is the model, not the people.

  3. 03

    Nobody owns the definition

    Active customer means three different things in three teams. Naming that, in public, is often the highest-value hour of the engagement.

  4. 04

    The report exists; the decision does not

    Reporting built without a decision attached becomes wallpaper. We start from the choice being made, and stop when it can be made.

E — What we build4 shapes of engagement
  1. D1

    Ingestion you can trust

    Scheduled extraction from the systems you already run, with schema changes surfaced rather than silently absorbed.

    • Source connectors
    • Incremental loads
    • Schema drift alerts
    • Replay
  2. D2

    Warehouse modelling

    Entities, grain and history modelled deliberately, so a join cannot quietly double a number.

    • Dimensional models
    • Slowly changing history
    • Tests on transforms
    • Documentation
  3. D3

    A semantic layer

    Metric definitions written once and reused everywhere, so every tool asking the same question gets the same answer.

    • Metric definitions
    • Ownership
    • Version control
    • Reconciliation checks
  4. D4

    Reporting people use

    Fewer dashboards, each answering a named question, with the definition and the freshness visible on the page.

    • Decision-first layout
    • Freshness indicators
    • Drill-through
    • Access control
F — ApproachHow the work runs
  1. 01

    Start from the decision

    Which choice is being made, by whom, how often. Everything downstream is derived from that answer.

  2. 02

    Define the metrics in writing

    Including the edge cases: refunds, cancellations, timezones, the customer who is also a supplier.

  3. 03

    Model, then move

    The shape of the warehouse is settled before pipelines are built, so pipelines are not rewritten to fit a late model.

  4. 04

    Reconcile out loud

    New numbers are compared against the old ones, and every difference is explained before the old report is retired.

L1Storage
PostgreSQLWarehousesColumnar storesPartitioning
L2Movement
ELT pipelinesBatch and incrementalOrchestrationReplay
L3Modelling
Dimensional designSQL transformsTestsLineage
L4Delivery
Semantic layersReportingEmbedded analyticsExports
G — Contact06 Clarity · resolved

Bring the two reports that disagree.

Send the numbers nobody can reconcile, and the decision they are supposed to support. We will find where the definitions diverge and what it would take to have one answer.

Send the two reportsApplied AI