This project was created as a solution to the final evaluation of OCTO Academy's AI Certification ("Développement et conduite de projets d'intelligence artificielle").

During the final evaluation, the candidate is given access to several raw files and SQLite DBs and asked to create a data science solution that can predict flights' delays based on historical data from several airports. We expect the candidate not to focus on creating a performing ML model (the data is fake and doesn't contain signal), but rather to propose an elegant end-to-end solution that could be deployed in production.

Disclaimer: As no additional data source was provided for the prediction part, the current prediction workflow simulates the existence of such data by predicting flight delays every 15min based on a random batch of 50 examples extracted from the training data. Of course, it wouldn't be the case in a real-world project!

The project structure served as an inspiration to the cookiecutter data science industrialization project template also available on my GitHub.

Create your local environment and install the project's dependencies (in dev mode):

conda create --name octo_fly python==3.7
conda activate octo_fly
pip install -e .

Train (or retrain) the model - this can take a few minutes:

python pipeline/train.py

Launch the Streamlit app and the orchestrated batch prediction service:

docker-compose -f docker/docker-compose-dev.yml up

Et voilà! You can check the web app at http://localhost:8501. The results will be updated every 15min with the last predictions.