Prediction MLOps Pipeline

The focus of this project was to get a hands-on experience with the full MLOps workflow. I used a common public prediction dataset, and I worked on the system around the model: training, tracking, model selection, deployment, and exposing predictions through an API.

• Designed and implemented an end-to-end prediction pipeline using Python, MLflow, FastAPI, and Docker.
• Built modular components for training, feature engineering, model evaluation, model comparison, and prediction.
• Used MLflow for experiment tracking, model artifacts, and model registry-style workflows.
• Deployed selected models through a FastAPI prediction service with request/response schemas and prediction tests.
• Containerized the MLflow tracking server, training workflow, and inference API with Docker Compose for reproducible local development and API testing.

The project is structured around three stages, wanting to illustrate how a data scientist would go from initial research, model building, to productionizing the model:

01_initial_ml_build/  -> initial notebook exploration
02_model_training/   -> training pipeline, feature code, model comparison, Airflow DAG
03_deployment/       -> FastAPI app, Dockerfile, bundled model, prediction tests

The API can serve either a selected MLflow run or a bundled local model. If a RUN_ID is provided, the service loads the model from MLflow:

runs:/<RUN_ID>/final_model

If no RUN_ID is set, it falls back to the local bundled model:

03_deployment/saved_models/model.joblib

That made the project easier to test locally while keeping the MLflow-backed workflow as the main model selection and serving path.

If I were to extend this project, I would use a more meaningful dataset and spend more time on the modeling process itself. For this version, the main goal was to understand the workflow end to end: how a model moves from training code, to tracked experiments, to a selected artifact, to an API that can serve predictions.