Case study / Finance operations control plane

FinanceOps Warehouse

A control-oriented system for moving provider data into client-isolated finance facts, close gates, evidence, work queues, and explicit human decisions.

The problem

Finance automation fails when it cannot explain its source, scope, or authority.

Provider feeds, workpapers, accounting policy, close status, and human approvals often drift into separate systems. The warehouse was designed to make those relationships explicit, client-scoped, and reviewable before an action reaches the books.

System Architecture

BigQuery is durable evidence and run truth. Repository state holds versioned client policy. The human control plane is a privacy-minimized release snapshot, not a substitute for the warehouse.

01 / Providers QBO / Stripe / Shopify / ShipBob / Bill.com Read-only ingestion by default

02 / Raw Client-isolated source records Source identity and freshness retained

03 / Canonical Normalized facts and controls All-or-rollback publication per refresh

04 / Agent views Close, FP&A, tax, evidence, JE review Explicit client and period scope

05 / Human action Review / approve / reopen / resolve Authority remains visible

What I built

Controls that fail loudly instead of manufacturing confidence.

Client isolationReusable workflows require an explicit client identity. Transaction rules do not cross clients without backtest and approval.

Close state machineLegal period transitions, compare-and-swap persistence, mandatory live gate reads, and an honest zero-review path.

Source trustFreshness, period coverage, evidence lineage, and missing-source states are separate from accounting conclusions.

Guarded postingOnly narrow QBO close mutations are allowed, with dry runs, environment and realm confirmation, idempotency checks, and post-write verification.

Operator releaseA human-facing portfolio shows status, source gaps, close posture, and next decisions without becoming financial truth.

Verification And Boundaries

The boundaries are part of the design, not footnotes.

Verified 838 automated tests passed

The full Django suite ran in this workspace on July 12, 2026, and the system check reported no issues.

Verified in repository records 22 warehouse quality checks

The project records a successful targeted canonical refresh with all quality checks passing.

Boundary A release snapshot is not live truth

The operator portfolio labels release and verified-live versions separately and fails closed when its snapshot is invalid.

Boundary Automation does not erase approval

Provider writes remain narrow and guarded. The system does not claim broad autonomous posting authority.

Why this belongs in a finance portfolio

It demonstrates the same judgment I bring to controllership: source discipline, explicit authority, reviewable decisions, and no quiet plugs.

Talk with me