GCP GenAI Engineer study plan
GCP GenAI Engineer · Give each week a clear focus.
8 weeks · About 8 hours per week
Developing ML models
Set up a Vertex AI Workbench notebook instance and run a custom training job using the Vertex AI SDK
Compare at least three model architectures (e.g., linear, deep neural network, gradient boosted trees) on a public dataset using Vertex AI Experiments
Explain how to handle overfitting and underfitting in the context of Vertex AI training, including regularization and early stopping
Implement a distributed training job using Vertex AI's reduction server or distributed training options for TensorFlow or PyTorch
Your checkpoint
A Jupyter notebook that trains and evaluates three models on a public dataset, logs metrics to Vertex AI Experiments, and includes a written summary of overfitting mitigation strategies.
Developing ML models & Designing data preparation and processing systems
Build a feature engineering pipeline using BigQuery and Dataflow to create a training dataset from raw data
Implement a custom model with a custom container using Vertex AI Training, including packaging and pushing the container to Artifact Registry
Design and implement a data validation step using TensorFlow Data Validation (TFDV) to detect anomalies and skew
Explain how to handle categorical and missing data in Vertex AI training, including preprocessing and feature crosses
Your checkpoint
A Dataflow pipeline that processes raw data into features, a custom training container that trains a model on those features, and a report from TFDV showing data statistics and anomalies.
Developing ML models & Serving and scaling ML models
Deploy a trained model to a Vertex AI Endpoint with automatic scaling and traffic split
Implement batch prediction using Vertex AI Batch Prediction and compare it with online prediction in terms of cost and latency
Explain how to optimize model serving for latency and throughput, including model optimization techniques like quantization and pruning
Set up a private endpoint and explain when to use private vs. public endpoints
Your checkpoint
A deployed model on Vertex AI Endpoint with a traffic split between two versions, plus a benchmark report comparing latency and cost of online vs. batch prediction.
Automating and orchestrating ML pipelines
Create a Vertex AI Pipeline that automates data preprocessing, training, and model deployment using the Kubeflow Pipelines SDK
Use Vertex AI Pipelines to schedule a recurring pipeline run and configure artifact lineage tracking
Implement conditional steps and caching in a pipeline to optimize execution time and cost
Explain how to integrate Cloud Build and Artifact Registry into a CI/CD pipeline for model training
Your checkpoint
A Vertex AI Pipeline that runs end-to-end on a schedule, with lineage tracking and caching enabled, and a diagram showing the pipeline steps and artifacts.
Automating and orchestrating ML pipelines & Monitoring, optimizing, and maintaining ML solutions
Set up Vertex AI Model Monitoring for a deployed model to detect training-serving skew and drift
Implement a retraining pipeline triggered by model monitoring alerts using Cloud Functions and Vertex AI Pipelines
Configure logging and auditing for Vertex AI resources using Cloud Logging and Cloud Audit Logs
Explain how to use Vertex AI Model Registry to manage model versions and metadata
Your checkpoint
A monitoring dashboard showing model performance metrics, an alert policy for skew detection, and a retraining pipeline that triggers automatically when an alert fires.
Architecting ML solutions
Design a machine learning solution architecture for a given business problem, including data sources, model training, and serving considerations
Compare and contrast Vertex AI services (AutoML, custom training, pipelines) and choose the appropriate one for a scenario
Estimate cost and performance trade-offs for different ML architectures (e.g., online vs. batch, real-time vs. micro-batch)
Explain security and compliance considerations for ML solutions on Google Cloud, including IAM, VPC Service Controls, and encryption
Your checkpoint
A written architecture design document for a hypothetical ML use case, including a diagram, service selection rationale, cost estimate, and security considerations.
Designing data preparation and processing systems & Monitoring, optimizing, and maintaining ML solutions
Implement a data pipeline using Cloud Data Fusion or Dataflow to handle streaming data ingestion and preprocessing
Use BigQuery ML to train and evaluate a model directly in BigQuery and export it to Vertex AI
Set up model explainability using Vertex AI's explanation features (e.g., Sampled Shapley) and interpret the results
Optimize model performance by analyzing monitoring data and applying techniques like hyperparameter tuning
Your checkpoint
A streaming data pipeline that feeds a BigQuery ML model, with explainability reports and a hyperparameter tuning job that improves model accuracy by at least 5%.
Architecting ML solutions & Developing ML models & Automating and orchestrating ML pipelines & Serving and scaling ML models & Monitoring, optimizing, and maintaining ML solutions
Complete a full practice exam under timed conditions and review incorrect answers to identify knowledge gaps
Review key concepts from all domains, focusing on weak areas identified in practice exams
Summarize best practices for each domain in a concise cheat sheet
Simulate a real-world scenario by designing and deploying an end-to-end ML solution that incorporates data preparation, training, deployment, and monitoring
Your checkpoint
A comprehensive study guide summarizing key concepts and best practices for all domains, plus a deployed end-to-end ML solution with documentation and monitoring enabled.
Milestones
Week 2: Successfully train a custom model using a custom container and demonstrate data validation with TFDV.
Week 4: Build and run a Vertex AI Pipeline that automates training and deployment, and show lineage tracking.
Week 6: Complete a practice exam and score at least 70% on questions from domains 1, 3, and 4.
Week 8: Achieve a score of 80% or higher on a full-length practice exam and deploy a complete ML solution with monitoring.