EllinCRM — Automação de Documentos com IA para Negócios Gregos
Chariton-kyp/EllinCRM
Extração de documentos com IA, busca híbrida grega e chat RAG com aprovação humana
Ver no GitHubAgent RAG baseado em LangGraph que extrai dados estruturados de formulários, e-mails e faturas com scores de confiança por campo, armazena com embeddings vetoriais para busca híbrida (semântica + keyword + trigram grego), e oferece chat IA em streaming com citações de fonte. Inclui workflow de aprovação humana, dashboard Next.js e sync automático para Google Sheets.
Instalação
npx convez-ai Chariton-kyp/EllinCRM1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Extrair dados de faturas e formulários com confiança por campo
- Buscar e recuperar registros via chat IA em grego com citações
- Implementar workflow de aprovação humana para OCR e extração de IA
- Sincronizar dados aprovados para Google Sheets automaticamente
Veredito

Referência sólida de engenharia de IA em produção: router multi-modelo, embeddings locais com pgvector, RAG híbrido para grego e UI polida. Code é source-available (read-only), ideal para aprender arquitetura enterprise de extração + chat com aprovação humana.
README
A tradução para o português deste README ainda está na fila.
EllinCRM
AI-powered document automation platform for Greek businesses — built to showcase production-grade AI engineering for automation services clients.
EllinCRM extracts structured data from forms, emails, and invoices using LLM-based extraction with per-field confidence scores, stores it with vector embeddings for hybrid semantic + keyword search, and provides a streaming AI chat agent that answers business questions in Greek with source citations — all under a human-in-the-loop approval workflow.

The records dashboard — forms, emails, and invoices side by side, with per-field confidence rings and one-click approve / edit / reject. Each extraction carries its source file and an audit trail.

The chat agent answering a Greek business question by calling structured tools, running hybrid retrieval over the records, and citing every source record it used. Tokens stream in live over SSE.
Highlights
- Multi-model LLM router — Gemini 3.1 Flash for extraction, Claude Sonnet 4.6 for chat, Claude Haiku 4.5 as fallback, all behind a single abstraction with automatic retry chains
- Local-first embeddings — EmbeddingGemma 300M runs on the server (768-dim, multilingual). Document content embeddings never leave your infrastructure
- Hybrid retrieval — pgvector cosine similarity (0.7) + PostgreSQL
tsvectorkeyword match (0.3), with Greek accent normalization and stopword filtering - Streaming RAG chat — SSE-based token streaming with in-line source citations (record ID + source file)
- Human-in-the-loop — Every extraction is reviewable, editable, approvable, and rejectable. Full audit trail in PostgreSQL
- Premium UI — Next.js 16 + React 19 + TailwindCSS v4 + shadcn/ui + Framer Motion. Dark mode, animated confidence rings, split-screen review, drag-and-drop upload, demo mode
- Google Sheets sync — Approved records auto-sync to Sheets with per-type tabs
Architecture
┌──────────────────────────────────────────────────────────────────┐
│ Next.js 16 Frontend │
│ Dashboard · Records · Extraction · Streaming Chat Widget │
└────────────────────────────┬─────────────────────────────────────┘
│ HTTP / SSE / WebSocket
┌────────────────────────────▼─────────────────────────────────────┐
│ FastAPI Backend (Python 3.13) │
│ ┌─────────────┐ ┌─────────────┐ ┌──────────────────────────┐ │
│ │ Extraction │ │ Records │ │ Chat (RAG) │ │
│ │ Router │ │ Router │ │ SSE Streaming Router │ │
│ └──────┬──────┘ └──────┬──────┘ └────────────┬─────────────┘ │
│ │ │ │ │
│ ┌──────▼────────────────▼───────────────────────▼─────────────┐ │
│ │ LiteLLM Router (Multi-Model Abstraction) │ │
│ │ gemini-flash · claude-sonnet · claude-haiku │ │
│ └─────────────────────────┬─────────────────────────────────────┘│
│ │ │
│ ┌─────────────────────────▼─────────────────────────────────────┐│
│ │ EmbeddingGemma 300M · Hybrid Search · Greek NLP ││
│ └─────────────────────────┬─────────────────────────────────────┘│
└────────────────────────────┼──────────────────────────────────────┘
│
┌──────────▼──────────┐
│ PostgreSQL 18 + │
│ pgvector │
└─────────────────────┘
Tech stack
| Layer | Technology |
|---|---|
| Backend | FastAPI (async), Python 3.13, SQLAlchemy 2.1, Pydantic 2.12, Alembic |
| Frontend | Next.js 16, React 19, TypeScript, TailwindCSS v4, shadcn/ui, Framer Motion |
| Database | PostgreSQL 18 with pgvector |
| Embeddings | EmbeddingGemma 300M via sentence-transformers (local, multilingual) |
| LLM Router | LiteLLM 1.81+ |
| LLMs | Gemini 3.1 Flash, Claude Sonnet 4.6, Claude Haiku 4.5 |
| State (FE) | TanStack Query 5 |
| i18n | next-intl (Greek + English) |
| Containers | Docker Compose v2 |
Getting started
Prerequisites
- Docker Desktop (or Docker Engine + Compose v2)
- A Google AI Studio API key (for Gemini)
- An Anthropic API key (for Claude)
- Optional: HuggingFace token for the gated
embeddinggemmamodel — falls back to a free multilingual model otherwise
1. Clone and configure
git clone https://github.com/Chariton-kyp/EllinCRM.git
cd EllinCRM
# Backend env
cp backend/.env.example backend/.env
# Edit backend/.env and add: GOOGLE_API_KEY, ANTHROPIC_API_KEY, HUGGINGFACE_TOKEN (optional)
# Frontend env
cp frontend/.env.example frontend/.env.local
2. Run (development with hot reload)
docker compose --profile dev up --build
| Service | URL |
|---|---|
| Frontend | http://localhost:7002 |
| Backend API | http://localhost:7000 |
| Backend docs | http://localhost:7000/docs |
| PostgreSQL | localhost:7001 |
3. Run tests
docker compose --profile test run --rm test
4. Production build
docker compose --profile prod up --build -d
Project layout
EllinCRM/
├── backend/ # FastAPI service
│ ├── app/
│ │ ├── ai/ # EmbeddingGemma, hybrid search, Greek NLP, LiteLLM router
│ │ ├── extractors/ # HTML/EML regex extractors (LLM fallback)
│ │ ├── routers/ # extraction · records · search · chat · notifications
│ │ ├── services/ # llm_extractor · rag_service · export · google_sheets
│ │ ├── db/ # SQLAlchemy models, repositories
│ │ └── core/ # config, logging, runtime settings
│ ├── alembic/ # Database migrations
│ └── tests/
├── frontend/ # Next.js 16 dashboard
│ ├── app/ # App Router pages
│ ├── components/ # UI: chat · dashboard · extraction · records · export
│ └── lib/ # API clients, hooks, utilities
├── dummy_data/ # 25 sample Greek business documents (forms, emails, invoices)
├── docs/ # Setup guides (Greek)
├── docker-compose.yml
└── README.md
Sample dummy data
EllinCRM ships with 25 realistic Greek business documents across 7 industries:
- Restaurant (Ταβέρνα "Ο Παράδεισος")
- Law firm (Δικηγορικό Γραφείο Παπαδόπουλος & Συνεργάτες)
- Accounting firm (Λογιστικό Γραφείο Οικονομίδης)
- Electronics retailer (TechnoShop)
- Construction company (ΕλληνΚατασκευές ΑΕ)
- Medical clinic (Ιατρικό Κέντρο "Υγεία")
- Hotel (Ξενοδοχείο "Αιγαίο")
Use the Demo Mode button on the dashboard to ingest all 25 in one click.
Privacy posture
- Embeddings are generated on-server by EmbeddingGemma — document content used for vector indexing never leaves your infrastructure
- Extraction and chat calls go to cloud LLMs (Gemini, Claude) — easily swappable for local models (Qwen, Mistral, Llama) on a sufficiently sized VPS
- Real
.envand credential files are gitignored by default — only.exampletemplates are tracked
License
Source-available for portfolio viewing only — not open source.
This repository is published publicly as a portfolio showcase of the author's work. You may read, study, and evaluate the code, but you may not use it in any product, service, or business — commercial or otherwise — and you may not redistribute, modify, or create derivative works without prior written permission.
See LICENSE.md for the full terms. For commercial licensing enquiries, contact haritos19@gmail.com.
Author
Chariton Kypraios — AI Engineer, Athens LinkedIn · EllinAI
🤖 Built with Claude Code