A/P Proactivity
Learns vendor cadence from 12 months of AP history, normalizes aliases, and flags expected-but-missing bills before close.
Output: prioritized review queue + Excel workpaper
Back to selected systems
Case study / Close operations platform
Howard Finance Platform
A three-part operating system for finding missing AP before close, reviewing GL coding with explainable suggestions, and routing owner statements through an attributed approval trail.
The problem
Close teams lose time in predictable places: missing bills, inconsistent coding, and approvals trapped in spreadsheets.
The platform makes those workflows visible before they become late-close surprises. It supports judgment with evidence and routing; it does not pretend that pattern matching is accounting authority.
Three Tools, One Operating Model
Each tool produces a reviewable queue or workpaper and preserves the identity of the client, period, source, run, and reviewer.
GL Coding Review
Uses vendor history and an explainable semantic layer to review coding, surface low-use accounts, and propose suggestions for ambiguous transactions.
Output: review/predict results + workbookOwner Statements
Routes generated statements through review, change requests, resubmission, approval, and payment with attributed transitions.
Output: status queue + audit trailArchitecture
Built as an operational product, not a spreadsheet macro.
FrontendNext.js 16, React 19, TypeScript, Tailwind, TanStack Table/Query, and Recharts.
BackendFastAPI, SQLAlchemy 2.0, Alembic, pandas, numpy, openpyxl, and Pydantic.
Data and accessPostgreSQL, client-scoped data paths, email/password authentication, and attributed workflow events.
Review engineA deterministic history layer handles the bulk; an explainable text layer handles ambiguous or novel cases without external model calls.
Verification And Boundaries
The July 10 retest covered the deployed GL review flow. It did not authorize or perform accounting-system posting.
Frontend verification
TypeScript and production build passed
Next.js compiled successfully and generated all 18 application routes in this workspace on July 12, 2026.
Live-app retest
Four client-scoped runs completed
Review and Predict modes were rerun across two isolated client scopes.
Isolation check
Zero cross-client suggestions
All four runs preserved account suggestions inside the selected client context.
Workpaper QA
Four workbooks opened cleanly
Expected sheets and identity lines were present, with zero formula-error matches.
Boundary
No QBO posting was performed
The tested writes were limited to the beta application: uploads, transaction ingest, run creation, and report downloads.
Boundary
Prediction inherits history
A known-good reference period is required. Low-confidence and novel cases remain human-review work.
Local verification limit
Backend tests were not rerun here
Docker is unavailable in this workspace; the public claim is therefore limited to the documented deployed-app retest.
Why this belongs in a finance portfolio