One Project To Learn MLOps

Let’s build a Machine Learning service to predict the Air Quality Index (AQI) in your city in the next 3 days, using a 100% serverless stack.

You will learn a lot, AND you will build something useful for society. Win-win

These are steps to build this ↓

Step 1 – Feature generation script

1 → fetches raw weather and pollutant data from an external API like https://aqicn.org

2 → computes features from this raw data (aka model inputs), and targets (aka model outputs)

3 → stores these features in the *Feature Store*

Step 2 – Backfill historical (features, targets)

To train a Machine Learning model later, you need enough historical data (features, targets) in your Feature Store. Run the feature script for a range of past dates, to get enough training data.

Step 3 – Model training script

1 → fetches historical (features, targets) from the Feature Store.

2 → trains and evaluates the best ML model possible for this data, e.g. XGBoostRegressor.

3 → stores the trained model in the Model Registry.

Step 4 – Automate execution of the feature script

Create a GitHub action to automatically run the feature script (from step 1) every hour.

GitHub actions are serverless computing power to run your code on a schedule. For free. Beautiful.

Step 6: Create a web app to show model predictions

Streamlit is a powerful Python library to develop and deploy web data apps.

Your app

1 → loads the model and features from the *Feature Store*

2 → computes model predictions and shows them on a beautiful UI.

BOOM!

Bonus

You can create another GitHub action to automate the model training script.

Why re-train the model?

Because ML model performance decreases over time. The best way to mitigate this is to regularly re-train the model, like once a week.

➡️ In this article you will find the recipe to implement automatic model re-training.