Hands-On Machine Learning Engineering & Operations

Design scalable machine learning pipelines by implementing feature engineering, an ingestion pipeline, and a centralized feature store with reproducible infrastructure, testing suites, and clear documentation.

Learn live feature engineering by building a feature store from raw data, converting features via an API, and speeding processing with multiprocessing and AsyncIO, plus feature expansion and unit testing.

Explore a refactored, object-oriented feature engineering pipeline with a chaining pattern, plus unit and integration testing, fixtures, and docker-based reproducible environments.

Refactor and version control feature store code by pushing to GitHub, then analyze data in BigQuery, visualize in Sheets, and apply Black, Pylint, Mypy, and profiling for MLOps workflows.

Explore machine learning engineering concepts from data quality and tuning to AutoML evaluation and scalable deployment, using a feature store in BigQuery and distributed training with Spark for at-scale modeling.

Explore live coding of model development from local runs to cloud-scale pipelines, comparing bagging, gradient boosting, and random forest, tuning hyper-parameters, and interpreting feature importance.

Engage in labs on Google Cloud Storage, Compute Engine, and Vertex AI to scale model development with blob storage, remote VMs, and automated feature-store workflows.

Master data design from bronze to gold, separate training and production pipelines, and manage the model lifecycle with MLflow to handle drift and leakage through champion-challenger and A/B testing.

Move from local notebooks to cloud-based model training using a feature store and an optimal model. Retrain on all available data to produce a production model with reproducible artifacts.

Leverage MLflow for model lifecycle management and experiment tracking, and use Spark on Dataproc for horizontal scaling, production deployment, and artifact packaging.

Design a production api to serve a trained model and compare dockerized deployment with serverless architecture. Monitor drift and access via the api using the feature store and BigQuery.

Develop and deploy a flask-based api for predicting win probabilities from a card hand, test with postman, and move from local serving to production via labs and google cloud functions.

Refactor notebook code into a deployable script and deploy the model with Google Cloud Functions, retrieving the production model from Google Cloud Storage and serving predictions via a post endpoint.

Explore experimentation frameworks for deploying machine learning models, comparing a/b testing, and contextual multi-armed bandits, and balancing exploration and exploitation to maximize real-world reward in production.

Explore experimentation in production machine learning, using ab testing and multi-armed bandits to compare models in real-world data, monitor drift, and optimize deployment via adaptive policies.

Learn to build a model drift pipeline with Airflow, using dockerized Airflow, a two-task etl process, feature store drift data, roc evaluation, and a weekly scheduled dag with alerts.

This final lecture highlights the key ideas in machine learning engineering, reviews the course, invites feedback, and encourages students to revisit code and pursue personal projects.