convez.ai
AgenteAutomaçãoAvançado
Chariton-kyp

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 GitHub
InicialPluginScore · de 0 a 100

Agent 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/EllinCRM

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 4 meses
Versão
v1.0.0
Issues abertas
15
Contribuidores
1

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.

Records dashboard with per-field confidence and human-in-the-loop review

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.

Streaming Greek-language RAG chat agent with citations

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 tsvector keyword 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

LayerTechnology
BackendFastAPI (async), Python 3.13, SQLAlchemy 2.1, Pydantic 2.12, Alembic
FrontendNext.js 16, React 19, TypeScript, TailwindCSS v4, shadcn/ui, Framer Motion
DatabasePostgreSQL 18 with pgvector
EmbeddingsEmbeddingGemma 300M via sentence-transformers (local, multilingual)
LLM RouterLiteLLM 1.81+
LLMsGemini 3.1 Flash, Claude Sonnet 4.6, Claude Haiku 4.5
State (FE)TanStack Query 5
i18nnext-intl (Greek + English)
ContainersDocker 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 embeddinggemma model — 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
ServiceURL
Frontendhttp://localhost:7002
Backend APIhttp://localhost:7000
Backend docshttp://localhost:7000/docs
PostgreSQLlocalhost: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 .env and credential files are gitignored by default — only .example templates 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

Ver o original no GitHub

Bom para

Relacionados