This repository contains an end-to-end ML engineering solution for the Microsoft Azure Predictive Maintenance dataset.

The goal is to predict whether a machine will fail in the near future, using a binary classification model.
The focus of this project is on reproducibility, deployment craftsmanship, and MLOps maturity, rather than state-of-the-art modeling.

git clone https://github.com/Maxkaizo/pred_maint.git
cd pred_maint

The project requires a .env file at the root directory.

If it does not exist, create it manually with the following content:

PREFECT_API_URL=http://prefect:4200/api
MLFLOW_TRACKING_URI=http://mlflow:5000
AWS_ACCESS_KEY_ID=test
AWS_SECRET_ACCESS_KEY=test
MLFLOW_S3_ENDPOINT_URL=http://localstack:4566
PYTHONPATH=/app
AWS_DEFAULT_REGION=us-east-1

This project is fully containerized. Running the following command spins up all required services:

docker compose up --build

This may take a while on the first run, as Docker will download images and initialize all services.

Services included:

• Prefect (pipeline orchestration) → UI at http://localhost:4200
• MLflow (experiment tracking & model registry) → UI at http://localhost:5000
• Postgres (metadata storage for Prefect & MLflow)
• LocalStack (S3 emulation for datalake & artifacts)
• Training pipeline (runs automatically on startup)
• Inference API (REST service for predictions at http://localhost:8000)

Once the containers are up, send a prediction request to the inference API:

curl -X POST http://localhost:8000/predict \
     -H "Content-Type: application/json" \
     -d @sample.json

.
├── app/                 # Core pipeline code (flows & tasks)
│   ├── flows/           # Prefect flows (orchestration entrypoints)
│   └── tasks/           # Modular tasks (data prep, FE, training)
├── data/                # Raw & processed data (local copy of Kaggle dataset)
├── docs/                # Documentation (TDD, design notes, tech report, diagrams)
├── inference_app/       # Inference service (REST API + model loader)
├── notebooks/           # EDA and experimentation notebooks
├── postgres-init/       # SQL init scripts for Postgres databases
├── docker-compose.yml   # Orchestration of all containers
├── Dockerfile*          # Container definitions (training, inference, mlflow)
├── environment.yml      # Python environment (dependencies)
└── sample.json          # Example payload for inference requests

• Training pipeline – Prefect orchestrates data ingestion, feature engineering, target creation, model training and registration in MLflow.
• Models – CatBoost (primary) and LightGBM (secondary), chosen based on PRC/ROC performance.
• Tracking – MLflow stores runs, metrics, and artifacts.
• Storage – LocalStack (S3 emulation) for datalake and artifacts.
• Deployment – Inference container serving the model via REST API.

• Best model: CatBoost (AP ≈ 0.86, F1 ≈ 0.87).
• Business trade-off: Recall prioritized (avoid missed failures) over precision (extra operational costs).
• Performance curves and full analysis can be found in docs/performance_curves.png.

• Technical Design Document (TDD)
• Design Notes (rationale & trade-offs)
• Short Tech Report
• System Architecture
• Walkthrough

Planned improvements (not included due to time constraints):

• Periodic retraining via Prefect schedules.
• Continuous monitoring with Evidently.
• CI/CD with GitHub Actions.
• Code linting & pre-commit hooks.