SimplifyEU
Administrative overhead for EU companies, handled by four small Python services instead of
one large application: duplicate invoices, forgotten subscriptions, a crew of AI
assistants, and the deadline reminders nobody wants to send.
🌐 Demo: simplify-eu.vercel.app — the marketing side is
open; the dashboard needs an account.
Table of contents
Overview
SimplifyEU targets the administrative work a small European company cannot avoid: invoicing
with the right VAT treatment, subscriptions that quietly renew, filings with deadlines, and
documents nobody has time to read.
The design choice that shapes the repository: one thin web surface, four independent
Python services. The Next.js app is the only thing a user talks to; each service owns one
problem, has its own Dockerfile, its own Supabase migration and its own API key. Nothing is
shared except types in packages/shared and the database.
Everything is localised in three languages (English, French, Spanish) with the locale in
the route segment.
The services
Web application — apps/web
The one surface a user sees: locale-prefixed routes, Supabase sign-in, and a dashboard that
fans out to whichever service answers the question.
- Next.js 16 App Router with
[locale] segments — every page exists per language by
construction
- Supabase SSR client — the session is read on the server, never trusted from the client
- Tailwind CSS 4 + Radix UI — mobile-first, with a bottom navigation bar on small screens
- Stripe for subscription billing, Recharts for the dashboard, next-themes for
dark mode, sonner for toasts, cmdk for the command palette
react-hook-form + Zod on every form; /ai-test verifies the AI integrations
Invoice duplicate detector — services/invoice-detector
The same invoice paid twice is a common and expensive accident. This service reads the PDFs,
falls back to OCR when there is no text layer, and matches on fuzzy similarity rather than
exact fields.
- pdfplumber + PyPDF2 for text and table extraction
- pytesseract + Pillow fallback, so a scanned invoice still gets read
- python-magic for real content-type detection on upload
- fuzzywuzzy + python-Levenshtein —
INV-2026/12 and INV 2026 12 are the same document
- slowapi rate limiting, so an upload endpoint cannot be used as a grinder
- Results stored as
invoices and duplicate_pairs
Subscription analyser — services/subscription-analyzer
Feed it a bank export and it reports what recurs, what rose in price and what nobody uses —
then sends the finding as a PDF.
- Celery workers, so a long analysis runs off the request path
- pandas for the tabular analysis; CSV, PDF (pdfplumber, pdf2image) and OCR
through the same pipeline
- PaddleOCR + paddlepaddle for scanned statements
- Recurrence classifier and price-rise detector kept deliberately separate
- Redis cache — re-running the same file is free
- reportlab / openpyxl to emit the PDF and spreadsheet reports
CrewAI assistants — services/crewai-agents
Five specialised agents rather than one prompt asked to be all five, each with its own
agent, crew and flow:
| Agent | Scope |
|---|
financial_assistant | Cash, invoices, financial insights |
tax_assistant | VAT and tax obligations |
document_manager | Document classification and linking |
operations_coordinator | Internal operations |
client_assistant | Client-facing answers |
- CrewAI + crewai-tools — roles, tasks and hand-off between agents
- FastAPI + gunicorn — each crew is reachable as a plain HTTP service
- APScheduler for recurring runs, Twilio for notifications, Redis for caching
- reportlab / openpyxl for generated documents
- API-key auth per service — the web app is the only legitimate caller
- Execution recorded in
agent_execution_logs, configurable via agent_settings
Team reminder bot — services/team-reminder-bot
Deadlines a small company keeps missing — filings, renewals, payments — chased automatically
instead of depending on who remembers.
- Teams and reminders API: who is reminded, of what, and how often
- dateparser + python-dateutil for human deadline expressions
- Supabase as the store, with its own migration
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
