Translatica is a production-ready AI-powered English → Spanish literary translation system designed to preserve tone, context, and narrative style. It leverages a LoRA-fine-tuned transformer (PEFT) to deliver high-quality translations with low inference cost, supported by a modular NLP pipeline, BLEU-based evaluation, and a clean full-stack web interface.

The system is Dockerized and deployment-ready, and can be scaled as a SaaS product for publishers and content platforms—demonstrating strong expertise in model optimization, end-to-end system design, and business-oriented AI engineering.

Try the real-time English → Spanish Translator here:
Translatica– Click Here

Bilingual-Bridge/
│
├── .github/
│   └── workflows/
│       └── main.yml
│
├── fine-tuned-model/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   └── README.md
│
├── fine-tuned-tokenizer/
│   ├── source.spm
│   ├── special_tokens_map.json
│   ├── target.spm
│   ├── tokenizer_config.json
│   └── vocab.json
│
├── logs/
│
├── notebook/
│   └── Experiment.ipynb
│
├── src/
│   ├── data.py
│   ├── logger.py
│   └── model.py
│
├── static/
│   ├── css/
│   │   └── style.css
│   └── js/
│       └── script.js
│
├── templates/
│   └── index.html
│
├── tests/
│
├── .gitignore
├── app.png
├── demo.mp4
├── Dockerfile
├── LICENSE
├── README.md
├── requirements.txt
├── setup.py
└── train.py

• Dataset: opus_books

  • English-Spanish parallel corpus
  • High-quality literary translation data

• Base Model: Helsinki-NLP/opus-mt-en-es

  • Pretrained NMT model for English → Spanish

• Tokenizer:

  • SentencePiece tokenizer (AutoTokenizer) from the base model

• Fine-Tuning Strategy:

  • PEFT (Parameter-Efficient Fine-Tuning) using LoRA (Low-Rank Adaptation)
  • Frozen base model — only LoRA-injected layers are trainable

• LoRA Configuration:

  • r = 8, alpha = 32, dropout = 0.1
  • Target Modules: ["q_proj", "v_proj"] (attention heads)

• Trainer Configuration:

  • Epochs: 3
  • Batch Size: 16
  • Learning Rate: 2e-5
  • Weight Decay: 0.01
  • Evaluation: Per Epoch
  • Early Stopping: Enabled
  • Mixed Precision: Enabled (fp16)

• Evaluation Metric:

  • BLEU Score using Hugging Face evaluate library

• Training Performance:

• Saving Checkpoints:

  • Tokenizer: tokenizer.save_pretrained("saved_tokenizer")
  • LoRA Adapter: peft_model.save_pretrained("saved_peft_model")

• Trainable Parameters:

  • Only ~0.38% of parameters were updated
  • Verified using: peft_model.print_trainable_parameters()

Observations:

• Training and validation loss gradually decreased.
• BLEU score stabilized over epochs, indicating early convergence.
• Indicates the model is learning but BLEU can be improved further with more epochs or tuning.

sequenceDiagram
    participant User
    participant WebUI as Web UI (Flask + HTML/CSS)
    participant TranslationServer as Translation App (Python)
    participant HFModel as Hugging Face Model

    User->>WebUI: Enter English text
    WebUI->>TranslationServer: translate(input_text)
    TranslationServer->>HFModel: Tokenize input
    HFModel-->>TranslationServer: Return tokenized input
    TranslationServer->>HFModel: Generate translation tokens
    HFModel-->>TranslationServer: Return translated tokens
    TranslationServer-->>WebUI: Return Spanish text
    WebUI->>User: Display translated output

FROM python:3.10
WORKDIR /app
COPY . .
RUN pip install -r requirements.txt
CMD ["python", "app.py"]

Build & run:

docker build -t translatica .
docker run -p 8501:8501 translatica

• Add multilingual support (more language pairs)
• Quantize and optimize for edge devices

git clone https://github.com/Md-Emon-Hasan/Translatica
cd translatica
pip install -r requirements.txt
streamlit run app.py

Md Emon Hasan
Email: [email protected]
WhatsApp: +8801834363533
GitHub: Md-Emon-Hasan
LinkedIn: Md Emon Hasan
Facebook: Md Emon Hasan