🌍 Universal AI Translator API (NLLB-200)

A high-performance translation API supporting 200+ languages using Meta's NLLB-200 model. Features optimized INT8 inference (600MB), event-driven architecture with Celery/Redis, and production-ready Docker deployment.

---

✨ Key Features

• ⚡ Fast & Lightweight: CTranslate2 INT8 quantization (4x faster, 2.5GB → 600MB)
• 🔄 Asynchronous Processing: Celery + Redis for non-blocking operations
• 🚀 Production Ready: Dockerized, Swagger docs, zero cold-start with singleton pattern
• 🌐 200+ Languages: English, Arabic, French, Chinese, Spanish, and more

---

🚀 Quick Start

Option 1: Docker Hub (Fastest - Recommended)

Run the pre-built containers directly from Docker Hub:

# Clone the repo
git clone https://github.com/Mohammed2372/Translation-API.git
cd Translation-API
# Pull and run the complete stack
docker-compose -f docker-compose.prod.yml up

This will automatically pull:

• mohammed237/translation-api:webv1 (Django API)
• mohammed237/translationapi:workerv1 (Celery worker with model)
• redis:7-alpine (official Redis image)

Access API:http://localhost:8000/api/docs/

Option 2: Build from Source

Prerequisites

• Docker & Docker Compose OR Python 3.10+ & Redis

Step 1: Get the Model

Option A - Use Kaggle Notebook (5 min):

1. Open: Model Preparation Notebook
2. Click "Copy and Edit" → Run all cells
3. Download translator_model.zip

Option B - Run Locally:

jupyter notebook notebooks/quantize_translator_model.ipynb
# test model
python test_translator_model.py

Step 2: Extract Model

/translationapi
├── model/
│ └── translator_model/ <-- Extract here
│ ├── model.bin
│ ├── config.json
│ └── shared_vocabulary.json

Step 3: Run

# With Docker
docker-compose up --build
# OR Manually (2 terminals)# Terminal 1
python manage.py runserver
# Terminal 2
celery -A core worker -l info # Add -P solo on Windows

---

📖 API Usage

Endpoint:POST /api/translate/

Request:

{
"text": "Hello, how are you?",
"source": "eng_Latn",
"target": "arb_Arab"
}

Response:

{
"status": "completed",
"result": {
"translated": "مرحبا، كيف حالك؟",
"original": "Hello, how are you?"
}
}

Interactive Docs:http://localhost:8000/api/docs/

---

🏗️ Architecture

graph LR
A[Client] -->|HTTP| B[Django API]
B -->|Task| C[Redis Queue]
C --> D[Celery Worker]
D -->|INT8 Model| E[CTranslate2]

Loading

Components:

• Django API: Request validation, task dispatch
• Redis: Message queue, traffic buffering
• Celery Worker: Background AI processing (model loaded once)
• CTranslate2: Optimized INT8 inference engine

---

🧠 Technical Highlights

Why CTranslate2?

Standard PyTorch is slow and heavy. CTranslate2 provides:

• 4x faster inference via C++ optimization
• INT8 quantization: 2.5GB → 600MB
• CPU-friendly (no GPU needed)

Singleton Pattern

Model loads once at worker startup (translationapi/ai_loader.py), eliminating 5+ second cold starts per request.

Smart Polling

API waits up to 10 seconds for results (feels real-time), then gracefully returns task ID for async checks under heavy load.

---

📁 Project Structure

translationapi/
├── translationapi/
│ ├── views.py # API endpoints
│ ├── tasks.py # Celery tasks
│ └── ai_loader.py # Model singleton
├── model/
│ └── translator_model/ # INT8 model (600MB)
├── notebooks/
│ ├── quantize_translator_model.ipynb
│ └── test_translator_model.py # test model if working
├── docker-compose.yml # Build images
├── docker-compose.prod.yml # Production with Docker Hub images
└── requirements.txt

---

📊 Performance

Metric	Value
Model Size	600MB (INT8)
Inference	~200ms/sentence (CPU)
Memory	~800MB total
Throughput	100+ req/min

---

📧 Links

• Kaggle Notebook:quantize-translator-model
• Docker Hub:
  • mohammed237/translationapi
• Issues:Report Bug