Unlike typical ML Demos or AI Wrappers, where you see calling an OpenAI/Gemini API and showing the respose, or coming up with just another prediction model having some fancy accuracy, - this project stands out to mimic real-world production grade workflows. 🔥✅
"The real engineering, is not in the model performance but its in the overall system's performance" is what being showcased. It covers the complete lifecycle of an ML solution from developing core ML pipeline, implementation of MLOps, creating inference API, which includes automated testing, deployment and serving with CI/CD - following industry best practices for scalibility, reproducibility and maintainablity.
📊 ML Pipeline: Fully modular and reproducible ML pipeline designed in a clean, scalable structure. Used DVC for data and pipeline tracking & version control.
⚙️ MLOps Tools: Experiments tracking, model versioning and registration with Mlflow. Automated testing with Tox configuration, Github Actions for CI/CD pipeline automation and deployment, Evidently for checking data drift, etc ..
⚡ Model Serving: Model inferencing through RESTful API for high performance and real-time predictions. Deployed on HuggingFace Spaces for accessible model serving.
📌 Note: Used LIME for model observability, which increased inference time to ~ 4.5secs when deployed! Performed optimization and redeployed, reducing inference time to ~ 0.3-0.6 secs. Still it can be have latency issues and cold start delays. All open source tools and free services have been used to make this application!
💡 For more info - please visit Github ↗️
💖 THANK YOU : with love and regards from Subinoy (developer) 🙏