Sentiment Analysis Microservice

This project is an end-to-end, containerized microservice for binary sentiment analysis. It consists of a Python backend that serves a Hugging Face Transformer model, a standalone script for fine-tuning, and a minimal React frontend to display the results.

---

Video demonstration: https://youtu.be/1_EXuoJUsjc

Tech Stack

• Backend: Python, FastAPI, Hugging Face transformers, PyTorch
• Frontend: React (with Vite), TailwindCSS, ShadCN, Nginx (for serving production build)
• Containerization: Docker, Docker Compose

---

Setup & Run Instructions

Prerequisites

• Docker
• Docker Compose

1. (Optional) Fine-Tuning the Model

The application can load a pre-trained model by default. However, you can fine-tune it on your own data using the provided script.

1. Prepare your data in a data.jsonl file inside the ./backend directory. Each line should be a JSON object, for example:

   {"text": "This is a great product!", "label": "positive"}

2. Navigate to the backend directory and create a virtual env then run the fine-tuning script. You can adjust the epochs and learning rate as needed.

   cd backend
   python -m venv myenv
   source myenv/bin/activate
   pip install -r requirements.txt
   python finetune.py --data data.jsonl --epochs 3 --lr 3e-5

   This saves the updated model weights to the ./backend/model directory, which will be automatically loaded by the API on the next startup.

2. Running the Application

The entire project can be built and run with a single command from the root directory.

docker compose up --build

Design Decisions

REST API over GraphQL

• The project uses a standard REST API.
• Reasoning:
  REST is simpler to implement for a single-purpose application and is widely understood by developers.
• Since the service provides just one prediction endpoint, using GraphQL would have added unnecessary complexity.

---

FastAPI as the Backend Framework

• The backend is built using FastAPI, a modern and high-performance web framework for Python.
• Key Benefits:
  • Speed: Extremely fast, thanks to asynchronous support.
  • Interactive Documentation: Automatically generates Swagger UI and OpenAPI docs.
  • Pydantic Integration: Simplifies request and response data validation using Python type hints.

---

Standard Hugging Face Model as Default

• The default model used is:
  distilbert-base-uncased-finetuned-sst-2-english
• Why this model?:
  • Balances speed and accuracy well.
  • Ideal for sentiment analysis.
  • Works out of the box without requiring fine-tuning.

---

Separation of Training and Inference

• The fine-tuning script is a separate command-line tool.
• Purpose:
  Keeps the API lightweight by offloading heavy training logic.
• Inference Only:
  • The API handles only prediction (inference).
  • Once training is complete, the API loads the trained model files during startup.

---

🐳 Docker Compose for Simplicity

• The project uses Docker Compose to manage services.
• Why Docker Compose?:
  • Manages both backend (FastAPI) and frontend (React) together.
  • Single-command startup:

    docker compose up --build

  • Simplifies local development and deployment.

---

This architecture ensures the application is modular, scalable, and easy to set up and use.

API Documentation

Interactive API documentation is automatically generated by FastAPI and is available at:

http://localhost:8000/docs

---

📍 Predict Endpoint

• URL: /predict
• Method: POST

📤 Request Body (JSON)

{
  "text": "Your text for analysis here"
}

Success Response (200 OK)

{
    "label": "positive",
    "score": 0.9998
}

The label will be either "positive" or "negative".