Skip to content
Taveras Solutions · Go to home
Book a diagnostic

MITAI invoicing

InvoiceFlow

AI invoice processing with DGII tax validation for the Dominican Republic.

InvoiceFlow turns a photo of an invoice into accounting-ready data and flags what needs a second look using tax rules. Invoices can arrive through the web or WhatsApp and go out to the DGII 606 report or an accounting system.

Stars
95
Forks
47
Language
Python / FastAPI
License
MIT
Detail of a demo invoice in InvoiceFlow: audit notes (16% ITBIS, invoice older than 90 days, missing 606 goods and services type), totals with ITBIS, the original image, and the accounting entry preview
95
GitHub stars
47
forks: teams that took the code as a base
9
export formats, including the DGII 606
44
API routes and screens in a single FastAPI service

The challenge

In the Dominican Republic every supplier invoice is typed in by hand: name, RNC, NCF, date, base amount, ITBIS, and withholdings. It is slow work, and one wrong digit in the RNC or NCF shows up later as an error in the 606 report filed with the DGII.

Invoices also do not arrive in one place. Many are snapped by an employee with a phone out in the field and sent over WhatsApp, so the accountant gets loose photos that still have to be moved into the accounting system.

What we built

We built InvoiceFlow with FastAPI and SQLAlchemy. It takes images or PDFs through a web upload or an Evolution API webhook for WhatsApp, limited to authorized numbers. OpenAI reads the document and returns structured fields: supplier, RNC, NCF, dates, base amount, ITBIS, withholdings, and line items.

The AI does not have the final word. Tax rules written in code run on top of the extraction and flag what needs a human look: ITBIS outside 16% or 18%, invoices older than 90 days, a missing NCF or 606 goods and services type. Each invoice keeps its confidence score and its alerts in plain sight.

The output goes straight to accounting: the 606 is filled into the DGII's official Excel template, and there are exports to CSV, Excel, JSON, QuickBooks, Xero, Odoo, and Contaplus, plus outgoing webhooks. To keep the OpenAI bill in check, the system has a daily dollar cap and an hourly request cap.

What it looks like

InvoiceFlow control center with the processing queue, invoices processed today, extraction quality, average cost per document, and recent activity
The dashboard shows the queue, extraction quality, and what each document costs.
Invoice register with three invoices selected and the export menu open on the DGII 606 option
Pick the invoices and export them in bulk to whatever format accounting uses.
Contents of the exported 606 file: RNC, ID type, goods and services type, NCF, date, amounts, and ITBIS for three demo invoices
The 606 comes out on the DGII's official template. A partial view of the columns, with the values from the real file.
AI engine settings: hidden OpenAI key, selected model, and daily dollar limit
A daily dollar cap prevents surprises on the OpenAI bill.

WHO IT'S FOR

  • Dominican businesses that receive many supplier invoices and type them in by hand for the 606 today.
  • Accounting firms that handle several companies and want invoices to arrive over WhatsApp instead of as loose photos.
  • Developers who need a base with the RNC, NCF, and ITBIS rules already written to build their own product.

WHAT IT DOES

  • OpenAI extraction of supplier, RNC, NCF, ITBIS, withholdings, totals, and line items from images or PDFs.
  • Automatic audit: alerts for missing data, ITBIS outside 16% or 18%, and invoices older than 90 days.
  • WhatsApp intake through Evolution API, limited to authorized numbers and with a status reply.
  • Exports to the DGII 606 format and to CSV, Excel, JSON, QuickBooks, Xero, Odoo, and Contaplus.
  • OpenAI cost control with a daily dollar limit and an hourly request limit.
  • A dashboard with real-time WebSocket updates and outgoing webhooks to other systems.

How it's built

  1. Intake The invoice arrives by web upload or through the Evolution API WhatsApp webhook.
  2. Deduplication Redis drops repeated messages, applies per-IP rate limits, and caches stats and settings.
  3. AI extraction openai_service.py reads the image, structures the fields, and applies the DGII tax validations.
  4. Storage SQLAlchemy stores invoices, users, and organizations in SQLite or PostgreSQL.
  5. Output A WebSocket dashboard, outgoing webhooks, and export_service.py for the 606 and accounting formats.

Stack

FastAPIPythonOpenAIWhatsAppSQLAlchemyPostgreSQLRedisWebSocketsTailwindAlpine.js

Why we built it

It comes from a real pain for Dominican businesses: typing in invoices and staying DGII compliant. It shows how Taveras Solutions combines AI, local tax rules, and WhatsApp in one flow.

Getting started

pip install -r requirements.txt
cp .env.example .env  # OPENAI_API_KEY, SECRET_KEY and admin user
python check_db.py
python main.py  # open http://localhost:8000

WHAT WE LEARNED

  • AI extracts, but tax rules decide. Keeping the RNC, NCF, and ITBIS checks in code rather than in the prompt makes the alerts predictable and easy to audit.
  • Filling the DGII's official template works better than inventing a format: the accountant opens the file they already know and validates it with the usual tool.
  • A service that calls a paid API needs caps from day one. The daily dollar limit and the hourly request limit are part of the product, not an add-on.

Need something like this built for you?

We use these same tools in our client projects.