Agentic Accounting Runtime — an example

A working example of a governed AI layer for bookkeeping: it reads documents, proposes the accounting treatment, and posts it to a real ledger — without fabricating the numbers. The accounting engine computes the double-entry; a second engine cross-checks it. Built in a few hours by composing open-source software.

This is a reference example, not a finished product. It demonstrates the approach and a tested data foundation. See what a full deployment adds below.

The idea

The hardest data problem in AI accounting is ground truth: no public dataset pairs a source document with its correct bookkeeping entry (real ledgers are private). So we don't fabricate entries — we generate them from a real accounting engine and cross-validate across a second one:

• ERPNext (GPL-3.0) posts each invoice; its engine computes balanced GL Entry rows — correct by construction.
• Odoo (LGPL-3.0) independently posts the same invoices; agreement between two engines = strong evidence the postings are real accounting.

What's here

Path	What
sandbox/generate_ground_truth.py	Posts invoices through ERPNext across 8 scenarios (incl. tax, multi-currency, discount, closed-period refusals, tax-mismatch), reads the GL back, verifies debits==credits
sandbox/cross_validate_odoo.py	Replays a slice through Odoo and compares the double-entry
sandbox/normalize_documents.py	Maps real CORD/SROIE receipt images into the runtime's document schema
sandbox/*.compose.yml	Disposable ERPNext (:8100) + Odoo (:8101), pinned to digests
docs/MANUAL.md	How the AI Runtime Stack is used, task by task, with clear AI-does / you-do boundaries
data/PROVENANCE.md	Every dataset, source, and license

First run: 5,100 postings, 100% balanced, 300 correct period-closed refusals, 0 errors; Odoo cross-validation 50/50 agree. Sample output: data/ground-truth/sample.jsonl.

Quickstart

# 1. bring up the disposable engines
cd sandbox && docker compose -p erpgen up -d          # ERPNext on :8100 (Administrator / admin)
docker compose -p odoogen -f odoo/docker-compose.yml up -d   # Odoo on :8101

# 2. generate correct-by-construction posting pairs
./run_generator.sh 500                                 # -> data/ground-truth/gtc_pairs.jsonl

# 3. cross-validate against the second engine
N=50 python3 cross_validate_odoo.py                    # -> erpnext_vs_odoo.jsonl

Review console (deployable UI)

A FastAPI + MD3 dashboard over the generated data — the review surface an accountant would use: KPI tiles, a review queue, and per-invoice document → proposed posting → GL entries → verification with approve / amend / reject. Same reference pattern as the ReDevOps agentic-app fleet. Runs on the committed sample.jsonl out of the box (point DATA_FILE at a full run to see all of it).

cd ui && docker build -t accounting-ui . && docker run --rm -p 8300:8300 accounting-ui
# or: pip install -r ui/requirements.txt && python3 -m uvicorn app:app --port 8300  (from ui/)

Then open http://localhost:8300. Set ERPNEXT_FRONT_URL for the "Open in ERPNext" link.

Going further

This example is the foundation. A production deployment adds: production document extraction (real W-2s/invoices/receipts), the QuickBooks Online / Xero / Drake adapters, the console / review UI, human-in-the-loop workflows, multi-tenant auth and SOC 2, e-file/MeF where relevant, and hardening at scale. We'd love to help build that — redevops.io.

License

AGPL-3.0 with the section-13 network clause waived. Third-party components and data sources: LICENSES.md.