🚀 Sentiment Analysis MLOps Pipeline

An end-to-end MLOps project demonstrating model deployment, observability, and performance optimization using FastAPI, AWS S3, and GitHub Actions.

This repository shows how to take an NLP model from fine-tuning to scalable inference — with dynamic quantization and CI/CD automation for real-world readiness.

---

🧠 Project Overview

This project serves a fine-tuned DistilBERT sentiment classifier through a production-grade API. It includes model loading from AWS S3, optional quantized inference, structured logging, load testing, and CI/CD automation.

Quantization reduced average latency by ≈60%, proving the practical value of lightweight model optimization.

Mode	Avg Latency (ms)	P95 Latency (ms)	Improvement
Without Quantization	3274.15	5912.65	—
With Quantization	1302.57	2581.50	✅ ~60% faster

---

🧩 Tech Stack

Category	Tools Used
Modeling	Hugging Face Transformers (DistilBERT)
Serving	FastAPI, Uvicorn
Deployment	AWS EC2, S3
Automation	GitHub Actions (CI/CD)
Monitoring	Custom structured logging (logs/app.log)
Testing	Pytest
Load Testing	Async load simulator with aiohttp
Optimization	PyTorch Dynamic Quantization

---

⚙️ Key Features

• 🔁 Automated model fetch from S3 during startup
• ⚙️ Dynamic quantization toggle for faster CPU inference
• 📈 Structured request logging (latency, client IP, text length, sentiment)
• 🧪 Pytest-based CI pipeline for stability
• 🌐 FastAPI endpoint for real-time predictions
• 📊 Load simulator to measure performance under concurrent requests

---

🧰 API Usage

Health Check

GET /ping

✅ Response:

{"status": 200, "quantized": false}

Predict Endpoint

POST /predict

Body:

{
  "text": "The movie was absolutely fantastic!",
  "quantize": true
}

✅ Response:

{
  "sentiment": "positive",
  "latency_ms": "1320.45",
  "quantized": true
}

---

📦 Project Structure

sentiment-mlops/
├── .github/
│   └── workflows/
│       ├── ci.yml              # Continuous Integration: pytest, lint checks
│       └── cd.yml              # Continuous Deployment: deploy to EC2
│
├── app/
│   ├── serve.py                # FastAPI app for serving predictions
│   └── utils.py                # Model loading, quantization, and inference logic
│
├── logs/
│   ├── app.log                 # Application logs
│   ├── latency-stats.txt       # Performance summary
│   └── simulator.log           # Load test results
│
├── scripts/
│   ├── load_simulator.py       # Simulates concurrent requests for stress testing
│   └── analyze_simulator_logs.py # Parses and visualizes latency results
│
├── deploy.sh                   # Shell script for EC2 deployment
├── Makefile                    # Unified dev commands (test, run, deploy)
├── requirements.txt             # Project dependencies
├── test_setup.py                # Basic API and health-check tests
└── README.md                    # Documentation (you’re reading it!)

---

🔄 CI/CD Pipeline

GitHub Actions automates:

• ✅ Environment setup (Python + dependencies)
• ✅ Linting & testing via pytest
• ✅ Failure alerts on PRs
• ✅ (Optional) Deployment to EC2 after passing tests

This ensures your API is always tested before merging, just like real MLOps production pipelines.

---

📊 Load Testing

Run async load simulation to test stability under concurrent requests:

python scripts/load_simulator.py

Results are logged to logs/simulator.log and visualized with a latency-over-time plot.

---

⚡ Quantization Impact

Metric	Without Quantization	With Quantization
Min Latency (ms)	331.52	291.97
Avg Latency (ms)	3274.15	1302.57
P95 Latency (ms)	5912.65	2581.50
Max Latency (ms)	9143.29	4830.67

👉 Demonstrates how PyTorch dynamic quantization reduces model size and speeds up inference — essential for CPU-based deployments.

---

🧪 Run Tests

make test

or

PYTHONPATH=. pytest -v

---

☁️ Deployment

The app is designed for AWS EC2 deployment. Model files are fetched from S3 on first run and cached locally.

uvicorn app.serve:app --host 0.0.0.0 --port 8000

---

🧠 MLOps Pitch

This project demonstrates a production-ready NLP deployment pipeline with performance optimization and automation at its core. By integrating AWS, FastAPI, and GitHub Actions, it showcases how MLOps turns research models into reliable, scalable services — cutting latency by 60% through smart model quantization.

---

👨‍💻 Author

M. Farrukh Mehmood
📫 smfarrukhm@gmail.com

🔗 LinkedIn | GitHub