🚀 MachineInnovators Reputation Monitor (MLOps)

Welcome to the Reputation Monitoring System, an end-to-end MLOps project designed to analyze social media sentiment using state-of-the-art AI models. This repository demonstrates a Level 4 MLOps Maturity pipeline, featuring automated testing, continuous integration, drift detection, automated retraining triggers, and cloud deployment.

---

🏗️ Architecture Overview

The system is built using a microservices architecture orchestrated by Docker.

Component	Technology	Description
Model	Hugging Face (RoBERTa)	twitter-roberta-base-sentiment-latest for SOTA sentiment analysis
Serving	FastAPI	High-performance REST API for real-time inference
Containerization	Docker	Ensures reproducibility across environments
CI/CD	GitHub Actions	Automates Build, Test, and Deploy workflows
Monitoring	Prometheus & Grafana	Real-time metrics collection (Latency, Throughput)
Drift Detection	Python Script	Checks model accuracy and triggers retraining if performance drops
Deployment	Hugging Face Spaces	Docker-based deployment for public access

---

🛠️ Project Structure

reputation-monitoring/
├── .github/
│   └── workflows/          # CI/CD Pipelines (Build, Test, Deploy, Retrain)
├── monitoring/             # Configuration for Prometheus & Grafana
├── src/                    # Source Code
│   ├── main.py            # FastAPI Application Entry Point
│   ├── model.py           # AI Model Logic (RoBERTa Wrapper)
│   └── drift_detector.py  # Script for performance monitoring
├── tests/                  # Integration & Load Tests
├── Dockerfile              # Blueprint for the application container
├── docker-compose.yml      # Orchestration of API + Monitoring Stack
└── requirements.txt        # Python Dependencies

---

🚦 CI/CD Pipeline Logic

Every time code is pushed to the main branch, the MLOps End-to-End Pipeline executes the following steps:

1. Build & Setup: Builds the Docker containers and installs dependencies (fixing PyTorch CPU versions to save space)
2. Start Services: Launches the full stack (API, Prometheus, Grafana) using Docker Compose
3. Integration Testing: Runs pytest to verify API health and inference logic
4. Load Testing: Simulates real traffic (tests/load_test.py) to generate live metrics in Prometheus
5. Drift Detection: The drift_detector.py script analyzes performance. If accuracy < 0.80, it triggers the Automated Retraining Workflow via GitHub API
6. Deployment (CD): If all tests pass, the application is automatically deployed to Hugging Face Spaces

---

🖥️ How to Run Locally

You can run the entire system on your machine using Docker.

Prerequisites

• Docker & Docker Compose installed

Steps

1. Clone the repository:

   git clone https://github.com/YOUR_USERNAME/YOUR_REPO_NAME.git
   cd reputation-monitoring

2. Start the Stack:

   docker compose up --build

3. Access the Services:

   • API (Swagger UI): http://localhost:8000/docs
   • Grafana Dashboard: http://localhost:3000 (User: admin, Pass: admin)
   • Prometheus: http://localhost:9090

---

🛡️ Project Design Choices

Why RoBERTa instead of FastText?

While the project prompt mentioned FastText, the specific requirement linked to the cardiffnlp/twitter-roberta-base-sentiment-latest model. We chose RoBERTa (a Transformer model) because it significantly outperforms FastText on social media text by understanding context, sarcasm, and slang.

Why GitHub Actions for Retraining?

Instead of using heavy orchestrators like Airflow (which consume excessive resources for a single pipeline), we implemented an Event-Driven Architecture. The monitoring script uses the GitHub API to trigger a repository_dispatch event, launching a dedicated GitHub Action workflow for retraining. This "Serverless MLOps" approach saves resources and is cost-effective.

---

👩‍💻 Credits

Project developed for the ProfessionAI MLOps Course.