Google Cloud Professional Machine Learning Engineer Exam Prep

What's inside

1. Level 1: Exam craft: the guide, the sections and how to studyFree
   [Exam craft: all six sections] Learn how the course maps onto Google's Professional Machine Learning Engineer exam guide (as of June 1, 2026). The guide belongs to Google and is not reproduced here; you can read it free on Google's certification page. Check Google's exam page for the current length, price and languages. This course is independent of Google and not endorsed by it, and no pass is promised. The level also covers the six section weightings, the product renames (Gemini Enterprise Agent Platform, formerly Vertex AI; Agent Platform AutoML; Agent Platform Model Registry) and how to handle scenario 'best answer' questions.
2. Level 2: BigQuery ML: matching model types to business problemsFree
   [Section 1: Architecting low-code AI solutions, 1.1] Learn to build models in BigQuery ML for classification, regression, forecasting and clustering with CREATE MODEL, and to choose the model_type (for example LOGISTIC_REG, LINEAR_REG, BOOSTED_TREE_CLASSIFIER, ARIMA_PLUS, KMEANS) that fits the business problem and the data that is already in the warehouse.
3. Level 3: BigQuery ML feature engineering and selection
   [Section 1: Architecting low-code AI solutions, 1.1] Learn to do feature engineering and selection inside BigQuery ML: preprocessing in the TRANSFORM clause (scaling, bucketizing, one-hot and feature crosses) so the same transforms run at prediction time, and using feature-importance and correlation checks to decide which columns to keep.
4. Level 4: BigQuery ML predictions and fine-tuning Gemini in BigQuery
   [Section 1: Architecting low-code AI solutions, 1.1] Learn to generate predictions with BigQuery ML (ML.PREDICT, ML.FORECAST, ML.EXPLAIN_PREDICT), and to fine-tune Gemini models from BigQuery by creating a remote model over a Cloud resource connection, training it with supervised tuning on prompt and label columns, and calling it from SQL.
5. Level 5: Training models with Agent Platform AutoML
   [Section 1: Architecting low-code AI solutions, 1.1] Learn to train models with Agent Platform AutoML for tabular classification, regression and forecasting, and for image and text tasks. Covers creating managed datasets, setting the training budget and the optimisation objective, and reading evaluation results in the console and through the Python SDK.
6. Level 6: Model Garden selection and industry-specific AI APIs
   [Section 1: Architecting low-code AI solutions, 1.2] Learn to evaluate and select the right model from Gemini Enterprise Agent Platform Model Garden (first-party, open and partner models, and models as a service). Also learn when a ready-made API such as Document AI, Vision API or Translate API solves the problem without training anything.
7. Level 7: Gemini, Imagen and Veo solutions: tuning and optimising
   [Section 1: Architecting low-code AI solutions, 1.2] Learn to build and tune solutions with Gemini, Imagen, Veo and models as a service in Model Garden, with RAG and agents in outline only. Also learn to optimise Gemini-based applications for cost, latency and availability: choosing the model tier, context caching, batch versus online calls, provisioned throughput and regional or global endpoints.
8. Level 8: Organising and exploring tabular, text and image data
   [Section 2: Collaborating within and across teams to manage data and models, 2.1] Learn to organise and explore different data types (tabular in BigQuery; text and images in Cloud Storage with manifest or metadata tables) so that the same data can be used efficiently for experimenting, training and serving.
9. Level 9: Choosing a preprocessing tool by scale and complexity
   [Section 2: Collaborating within and across teams to manage data and models, 2.1] Learn to choose the right tool for data preprocessing based on scale and complexity: BigQuery SQL, Dataflow (Apache Beam), Apache Spark (Dataproc or serverless Spark) or in-memory Python frameworks such as pandas.
10. Level 10: Creating and consolidating features in Feature Store
    [Section 2: Collaborating within and across teams to manage data and models, 2.1] Learn to create and consolidate features in Agent Platform Feature Store with BigQuery as the offline source: feature groups, features, entity IDs and timestamps. Shows how teams share one governed set of features instead of each rebuilding its own.
11. Level 11: Data privacy and handling PII for ML
    [Section 2: Collaborating within and across teams to manage data and models, 2.1] Learn how to keep data private and handle sensitive information such as PII in ML datasets: discovering it and de-identifying it with Sensitive Data Protection (formerly Cloud DLP), masking, tokenising or bucketing, and controlling access with IAM, policy tags and column-level security in BigQuery.
12. Level 12: Notebook setup: collaboration and security
    [Section 2: Collaborating within and across teams to manage data and models, 2.2] Learn the collaboration and security best practices for setting up and running notebook environments in Agent Platform Workbench and Colab Enterprise: runtime templates, service accounts, VPC Service Controls, private IPs, CMEK, idle shutdown and sharing.
13. Level 13: Prototyping in notebooks with frameworks and Model Garden
    [Section 2: Collaborating within and across teams to manage data and models, 2.2] Learn to develop models in Agent Platform Workbench or Colab Enterprise notebooks with common frameworks (PyTorch, sklearn, JAX), and to prototype with foundation and open-source models from Model Garden inside the notebook.
14. Level 14: Experiments, metadata and lineage tracking
    [Section 2: Collaborating within and across teams to manage data and models, 2.3] Learn to choose the right development and experimentation environment for your framework: Experiments on Gemini Enterprise Agent Platform, Agent Platform Pipelines or Kubeflow Pipelines. Also learn to track and compare model artifacts, versions and lineage with Experiments and Agent Platform ML Metadata.
15. Level 15: Evaluating predictive and gen AI solutions
    [Section 2: Collaborating within and across teams to manage data and models, 2.3] Learn to evaluate predictive and gen AI solutions: classification, regression and ranking metrics; threshold choice; computed metrics for generated text; and LLM-as-a-judge with the platform's gen AI evaluation tooling, including pointwise and pairwise rubrics.
16. Level 16: Choosing the model type and interpretability approach
    [Section 3: Scaling prototypes into ML models, 3.1] Learn to choose the model type for a task (for example ARIMA, gradient-boosted trees, DNN or LLM) by trading off cost, complexity, latency and scalability, and to pick modelling techniques that meet interpretability requirements.
17. Level 17: Choosing the product and the deployment strategy
    [Section 3: Scaling prototypes into ML models, 3.1] Learn to choose the product (Agent Platform AutoML, BigQuery ML, custom training or Agent Platform Pipelines) and the deployment strategy (batch versus online, managed endpoint versus in-warehouse prediction) that fit the task's cost, complexity, latency and scalability limits.
18. Level 18: Organising training data in Cloud Storage and BigQuery
    [Section 3: Scaling prototypes into ML models, 3.2] Learn to organise training data of every type (tabular, text, speech, images and videos) on Google Cloud: when to use BigQuery and when to use Cloud Storage, bucket layout and naming, file formats, sharding, and keeping training, validation and test splits reproducible.
19. Level 19: Ingesting structured and unstructured data into training
    [Section 3: Scaling prototypes into ML models, 3.2] Learn to ingest structured and unstructured data from different sources (databases, streams, on-premises files, other clouds) into training pipelines, using tools such as BigQuery Data Transfer Service, Storage Transfer Service, Pub/Sub with Dataflow, Datastream, and the BigQuery Storage Read API for fast reads.
20. Level 20: Agent Platform custom training: containers and jobs
    [Section 3: Scaling prototypes into ML models, 3.2] Learn to train models with Agent Platform custom training: prebuilt training containers or custom containers, CustomTrainingJob and CustomJob in the Python SDK, worker pool specs, environment variables such as AIP_MODEL_DIR, and writing artifacts to Cloud Storage.
21. Level 21: Troubleshooting ML model training failures
    [Section 3: Scaling prototypes into ML models — 3.2 Training models] Learners work out why custom, AutoML and Kubeflow-on-GKE training jobs fail, using job logs in Cloud Logging, error states, quota and permission errors, out-of-memory crashes, data and path problems, and loss that turns into NaN or never drops.
22. Level 22: Hyperparameter tuning on Agent Platform
    [Section 3: Scaling prototypes into ML models — 3.2 Training models] Learners set up hyperparameter tuning jobs on Gemini Enterprise Agent Platform (formerly Vertex AI) with the Python SDK and the console: parameter specs, a metric to optimise, search algorithms, how many trials run in total and at once, and early stopping. They also see how this differs from tuning options in BigQuery ML and AutoML.
23. Level 23: Fine-tuning foundational models: when and how
    [Section 3: Scaling prototypes into ML models — 3.2 Training models] Learners decide when tuning a foundational model is worth it compared with prompt and context engineering or RAG. They then run supervised fine-tuning of Gemini on Agent Platform and parameter-efficient tuning (e.g. LoRA) of open models from Model Garden, including how to prepare the dataset and evaluate the tuned model.
24. Level 24: Choosing CPU, GPU or TPU for training
    [Section 3: Scaling prototypes into ML models — 3.3 Choosing appropriate hardware for training] Learners weigh CPU, GPU and TPU options for training against the model architecture, framework support, memory needs, cost and availability, and attach accelerators to Agent Platform custom training jobs.
25. Level 25: Distributed training: data and model parallelism
    [Section 3: Scaling prototypes into ML models — 3.3 Choosing appropriate hardware for training] Learners compare data parallelism with model parallelism on GPUs and TPUs, map them to framework tools such as PyTorch DDP/FSDP, TensorFlow distribution strategies and JAX sharding, and set up multi-worker pools on Agent Platform custom training.
26. Level 26: Batch vs online inference on Agent Platform
    [Section 4: Serving and scaling models — 4.1 Serving models] Learners choose between batch prediction jobs and online prediction on endpoints, based on latency, volume, freshness and cost. They run both with the Python SDK and the console, using BigQuery and Cloud Storage as inputs and outputs.
27. Level 27: Serving on Model Garden, Cloud Run and GKE
    [Section 4: Serving and scaling models — 4.1 Serving models] Learners compare serving options: one-click deployment from Model Garden, models as a service, Agent Platform endpoints, Cloud Run and GKE. They weigh how much each manages for you, its GPU support, how it scales to zero and the team skills it needs.
28. Level 28: Prebuilt and custom serving containers
    [Section 4: Serving and scaling models — 4.1 Serving models] Learners package PyTorch, XGBoost, scikit-learn and TensorFlow models for serving. They use prebuilt prediction containers where the framework is supported and write custom containers or Custom Prediction Routines where it isn't, meeting the platform's container requirements for health and predict routes.
29. Level 29: Organising models in Model Registry
    [Section 4: Serving and scaling models — 4.1 Serving models] Learners organise and version models in Gemini Enterprise Agent Platform Model Registry (formerly Vertex AI Model Registry): model versions, the default version and aliases, labels, registering BigQuery ML models, and deploying or evaluating a chosen version.
30. Level 30: Rollouts: A/B testing and canary deployments
    [Section 4: Serving and scaling models — 4.1 Serving models] Learners compare model versions by deploying several models to one endpoint with a traffic split, run canary and A/B rollouts, and plan rollback. They also contrast this with shadow-style comparison set up in your own serving code.
31. Level 31: Inference preprocessing and postprocessing
    [Section 4: Serving and scaling models — 4.1 Serving models] Learners design how requests are transformed before and after the model runs. Options include Custom Prediction Routines, preprocessing inside the model graph, and preprocessing in the calling application, plus postprocessing such as thresholds, label mapping and formatting responses.
32. Level 32: Serving features from Agent Platform Feature Store
    [Section 4: Serving and scaling models — 4.2 Scaling online model serving] Learners serve features at low latency from Agent Platform Feature Store: BigQuery as the source of truth, online store options, sync schedules and fetching features at request time to enrich prediction requests.
33. Level 33: Public and private endpoints
    [Section 4: Serving and scaling models — 4.2 Scaling online model serving] Learners deploy models to public endpoints, dedicated endpoints and private endpoints reached through Private Service Connect or VPC Network Peering, and match each to security, latency and networking requirements.
34. Level 34: Serving hardware and scaling the backend
    [Section 4: Serving and scaling models — 4.2 Scaling online model serving] Learners choose CPU, GPU, TPU or edge hardware for inference and scale the serving backend to match throughput. This covers minimum and maximum replica counts, autoscaling targets, Agent Platform Inference compared with containerised serving on GKE or Cloud Run, and edge deployment for offline or low-latency needs.
35. Level 35: Tuning models for production training and serving
    [Section 4: Serving and scaling models — 4.2 Scaling online model serving] Learners make models cheaper and faster in production with quantisation, distillation, pruning, graph compilation, request batching and right-sizing hardware, while checking that accuracy stays within an agreed budget.
36. Level 36: Validating data and models in pipelines
    [Section 5: Automating and orchestrating ML pipelines — 5.1 Developing end-to-end ML pipelines] Learners add validation gates to pipelines: schema and statistics checks on incoming data (e.g. TensorFlow Data Validation), checks for anomalies and skew, model evaluation against a baseline, and conditional deployment only when a model passes.
37. Level 37: Building Agent Platform Pipelines with KFP
    [Section 5: Automating and orchestrating ML pipelines — 5.1 Developing end-to-end ML pipelines] Learners write custom pipelines with the Kubeflow Pipelines (KFP) SDK, covering components, inputs and outputs, artifacts and caching. They compile pipelines and run them on Agent Platform Pipelines (formerly Vertex AI Pipelines) from the SDK and the console.
38. Level 38: Pipelines from templates and prebuilt components
    [Section 5: Automating and orchestrating ML pipelines — 5.1 Developing end-to-end ML pipelines] Learners speed up pipeline building with Google Cloud Pipeline Components, pipeline templates and Tabular Workflows, and decide when a template is enough and when you need custom code.
39. Level 39: Orchestrators: Airflow, Ray and Agent Platform
    [Section 5: Automating and orchestrating ML pipelines — 5.1 Developing end-to-end ML pipelines] Learners compare managed and unmanaged orchestration: Agent Platform Pipelines, Managed Service for Apache Airflow (formerly Cloud Composer), Ray on Gemini Enterprise Agent Platform and self-managed Kubeflow on GKE. They choose based on what the workload needs, e.g. cross-system data workflows, distributed Python compute or ML-native lineage.
40. Level 40: Consistent preprocessing for training and serving
    [Section 5: Automating and orchestrating ML pipelines — 5.1 Developing end-to-end ML pipelines] Learners prevent training-serving skew by keeping transformations identical in both paths. Approaches include TensorFlow Transform, preprocessing layers inside the model, shared preprocessing code packaged with the model, BigQuery ML TRANSFORM and features served from Feature Store.
41. Level 41: Setting a retraining policy
    [Section 5: Automating and orchestrating ML pipelines] Serves 5.2, determining an appropriate retraining policy: you compare retraining on a schedule, when new data arrives, when model performance drops, or when Model Monitoring on Agent Platform (formerly Vertex AI Model Monitoring) raises a drift alert. You then weigh each option's cost, data freshness and risk.
42. Level 42: CI for ML code and pipelines with Cloud Build
    [Section 5: Automating and orchestrating ML pipelines] Serves 5.2, deploying models in CI/CD/CT pipelines with Cloud Build: you build the continuous integration half. That means unit-testing components, building training and serving containers into Artifact Registry, and compiling and uploading Kubeflow Pipelines definitions whenever code changes.
43. Level 43: CD/CT: gated model promotion and rollback
    [Section 5: Automating and orchestrating ML pipelines] Serves 5.2 and completes CI/CD/CT. A continuous training pipeline evaluates each new model against the current champion, registers the winner as a new Model Registry version, and lets Cloud Build deploy it to an endpoint, with an approval step and a rollback path.
44. Level 44: Stopping data and model leaks
    [Section 6: Monitoring AI solutions] Serves 6.1, building secure AI systems: you protect against data exfiltration and against sharing sensitive data with LLMs. The tools are regex and Sensitive Data Protection redaction of prompts and responses, least-privilege IAM, and network perimeters around models and training data.
45. Level 45: Malicious prompting: safety filters and Model Armor
    [Section 6: Monitoring AI solutions] Serves 6.1's malicious-prompting item: you defend Gemini-based applications against prompt injection, jailbreaks and harmful output by layering Gemini safety filter settings with Model Armor screening of prompts and responses.
46. Level 46: Responsible AI and bias monitoring
    [Section 6: Monitoring AI solutions] Serves 6.1, aligning with responsible AI practices: you measure and monitor bias across data slices and sensitive groups, both before launch and in production. You also learn which mitigations to apply to data, model or threshold, and how to document decisions for other teams.
47. Level 47: Explainability on Agent Platform
    [Section 6: Monitoring AI solutions] Serves 6.1, model explainability on Agent Platform: you configure feature attributions (Sampled Shapley, Integrated Gradients, XRAI) and example-based explanations on models served through Agent Platform Inference, and read explanations returned with predictions.
48. Level 48: Model Monitoring: skew, drift and attribution drift
    [Section 6: Monitoring AI solutions] Serves 6.2: you configure Model Monitoring on Agent Platform (formerly Vertex AI Model Monitoring) to evaluate production models continuously. You learn to tell apart training-serving skew, data drift, concept drift and feature attribution drift, and to troubleshoot the alerts each one raises.
49. Level 49: Monitoring and evaluating gen AI solutions
    [Section 6: Monitoring AI solutions] Serves 6.2, monitoring, testing and evaluating gen AI solutions: you use Gen AI evals (formerly the Gen AI evaluation service on Vertex AI), with computed metrics and LLM-as-a-judge, for regression tests before release and for continuous evaluation in production. You also learn to trace and troubleshoot quality, latency and cost problems.
50. Level 50: Timed mock exam Timed mock
    A timed mock exam of 50 original scenario questions in 120 minutes, drawn from all six sections of the exam guide in proportion to their weights; the pass mark is 70%.

Access

The first 2 levels are free with a free account. Every level, the podcast edition and the AI tutor come with All Access at £4.99/month or any Creator plan — see pricing.

Start free →