Case — Dux NF (AI-Powered Invoice Automation via E-mail)

Type: Freelance (client: Dux Logistics — pickup operation at Fiat/Citroën dealerships, Stellantis context) Role: Full Stack Developer (full authorship) Status: In production Stack: Python, FastAPI, React + TypeScript + Vite, PostgreSQL, Microsoft Graph API (Outlook/M365), AI for extraction (OCR/LLM fallback), Docker Compose, deployment on Hostinger/Oracle Cloud

Context and problem

The logistics operation received dozens of invoices (XML and PDF) by e-mail and manually fed an operational demand-control spreadsheet used with Stellantis. The process was slow, prone to typing errors and lacked auditing — and the spreadsheet had formulas and validations that couldn't be broken.

Solution

An end-to-end pipeline that monitors the Outlook/Microsoft 365 mailbox via Microsoft Graph, captures invoice attachments, extracts the data and automatically fills in the operational spreadsheet — with an admin panel for operations, a failure queue and configuration.

Layered extraction strategy (cost and reliability):

NF-e XML — deterministic parsing (primary source, zero AI cost)
PDF text — deterministic parsing when possible
OCR/AI — only when XML is missing/invalid and the PDF isn't parseable

Operational dashboard with quality and cost breakdowns: total received, success via XML, success without AI, processed with AI/OCR, failures with reason and manual reprocessing action.

Architecture and technical decisions

FastAPI + a lightweight async worker (simple scheduling, no distributed queue in the MVP — a conscious decision for operational simplicity)
Microsoft Graph to read the monitored mailbox, with folder and parameters configurable from the frontend itself
Non-destructive spreadsheet writing: inserts rows without breaking formulas, validations or out-of-scope columns — including an audited backfill with a prior backup (backup_pre_backfill.sql)
Attachments archived in a controlled directory for auditing and reprocessing
PostgreSQL for a complete processing trail (what came in, how it was extracted, what failed and why)
Documented AI fallback plan (plano-fallback-ia-extracao-nf.md): AI as a last resort, with separate metrics on the dashboard to control cost
Secrets exclusively on the backend; the frontend never sees the token/client secret
React + TypeScript + Vite for the panel; Docker Compose with variants for Hostinger, Oracle and ngrok (demo)

Challenges and solutions

Reliability over "AI magic": a deterministic-first hierarchy made AI cost marginal and the success rate auditable per breakdown on the dashboard.
Live spreadsheet shared with the end client: surgical writes preserving the artifact the operation already used — adoption without a change in habits.
Operational business rules (dealership CNPJ → pickup point): implemented as automatic enrichment of the extracted data.

Results and impact

Eliminated manual invoice entry into the operational spreadsheet [volume/day TO CONFIRM]
Full auditing: every processed invoice has its origin, extraction method and result recorded
Controlled AI cost: most documents resolved via deterministic parsing [percentage TO CONFIRM on the dashboard]