Skip to main content

Proof capsule

Accounting Workflows

Document processing, semantic data models, and replayable runs

A governed Python pipeline turning document and source records into reproducible financial views and controlled reports.

Inspect workflow codeDocumentation repositoryBack to Engineering

Problem

Financial documents are difficult to reconcile when each report reinterprets source records. A useful workflow needs canonical representations, explicit semantic marts, exact-run replay, and a privacy boundary around source material.

What I built

I developed an inspectable pipeline for document/source ingestion, canonical materialization, debt and treasury views, reconciliation, governed reports, and report-bundle publication. Each stage can be replayed for an exact run scope, while public documentation uses sanitized evidence only.

My role and ownership

  • Designed canonical data representations and stage/run boundaries.
  • Implemented validation, semantic marts, report contracts, and exact-run replay paths.
  • Documented smoke and publication-dry-run checks while keeping private financial records out of public evidence.

Key decisions

Materialize canonical records before semantic marts or reports.

Tradeoff: Adds an intermediate model, but gives reconciliation a stable comparison point.

Scope every stage to an exact run and preserve replay metadata.

Tradeoff: Run bookkeeping is deliberate, but failures can be reproduced.

Keep public proof at contract/test/documentation level.

Tradeoff: Less visual spectacle, but avoids leaking private data or implying certification.

Reliability & testing

The repository defines smoke, report-contract, docs-input-bridge, and run-scope tests. The audit did not observe them passing, so they are defined checks rather than production certification.

  • test_governed_metrics_pipeline_smoke.py: pipeline smoke path.
  • test_reports_contracts.py: governed report contracts.
  • test_docs_input_bridge.py and test_run_scope_contract.py: input and exact-run boundaries.

Recovery: Exact-run replay and validation provide a controlled route to reproduce a failed stage.

What this proves

  • Design data workflows around canonical representations and semantic contracts.
  • Make financial/document processing replayable and reconcilable by run scope.
  • Communicate a privacy boundary instead of using private records as portfolio evidence.