Skip to content
← All case studies
Finance reporting automation Capability project

Controlled workbook-to-board-pack pipeline

Built an eight-stage pipeline that turns validated workbooks into self-contained board packs, passing 71 of 71 pipeline tests and 15 of 15 business gates.

71 / 71
Pipeline tests
15 / 15
Business gates
~58-page offline PDF
Output

Evidence inputs

  • A stack-agnostic data contract feeding an offline, confidential-by-design pack generator.
  • Python · Pure-Python HTML/SVG renderer · YAML/JSON contracts

Transformation

  • Built an eight-stage pipeline behind a reusable data contract.
  • Enforced control-total, source-freshness, offline-rendering and reader-comprehension gates.
  • Restricted model judgement to validated data; the model cannot create figures.

Controls & assurance

  • Failed closed when source evidence was stale or unverified.
  • Required extracted figures to equal the source statement, preventing internally consistent but incorrect packs.

Output

  • Passes 71 of 71 pipeline tests and 15 of 15 business gates, producing a self-contained board pack of roughly fifty-eight pages with no external dependencies.

Business value

  • Gives executives a decision pack with reconciliation and comprehension enforced by controls.
  1. Problem

    Manual pack production did not guarantee source reconciliation, fresh evidence or clear interpretation by non-technical readers.

  2. Approach

    Built an eight-stage pipeline behind a reusable data contract.

  3. Outcome

    Passes 71 of 71 pipeline tests and 15 of 15 business gates, producing a self-contained board pack of roughly fifty-eight pages with no external dependencies.

Business value

Gives executives a decision pack with reconciliation and comprehension enforced by controls.

Transformation route

  1. 01

    Built an eight-stage pipeline behind a reusable data contract.

  2. 02

    Enforced control-total, source-freshness, offline-rendering and reader-comprehension gates.

  3. 03

    Restricted model judgement to validated data; the model cannot create figures.

Decision log

  • Failed closed when source evidence was stale or unverified.
  • Required extracted figures to equal the source statement, preventing internally consistent but incorrect packs.

What I learned

  • A real-reader comprehension gate catches decision failures that arithmetic checks cannot.

Future improvements

  • Add further pack types and widen the external-source reconciliation anchors.

← All case studies · Discuss this work →