PMLE syllabus: every exam domain and topic

Architecting low-code AI solutions

13%

Solving problems with the least effort: BigQuery ML, pre-trained APIs, foundation models, RAG and AutoML. About 13% of the exam.

Developing models with BigQuery ML

Choosing the model type for the business problem: linear and logistic regression, boosted trees, time series, matrix factorisation, k-means, autoencoders
Feature engineering and selection in BigQuery ML: TRANSFORM, preprocessing functions, regularisation, feature importance
Generating predictions with ML.PREDICT, ML.FORECAST and ML.RECOMMEND, and serving through Vertex AI Model Registry

ML APIs and foundation models

Building applications from ML APIs and Model Garden models, and choosing managed API versus self-deployed
Industry-specific APIs such as Document AI and the Retail API
Retrieval augmented generation with Vertex AI Agent Builder, Vertex AI Search and RAG Engine

Training models with AutoML

14%

Managing data, features, notebooks and experiments so teams can share and reproduce work. About 14% of the exam.

18%

Turning a notebook model into a robust, trainable, tuned and well-provisioned one. About 18% of the exam.

20%

Getting predictions to users reliably, at the right latency and cost. About 20% of the exam.

Serving models

Scaling online model serving

Vertex AI Feature Store lookups at serving time
Public and private endpoints
Choosing CPU, GPU, TPU or edge hardware for serving
Scaling the serving backend on throughput, with Vertex AI Prediction or containerised serving
Tuning models for production: simplification, quantisation, latency, memory and throughput

22%

Making the ML lifecycle repeatable: pipelines, retraining, CI/CD and lineage. About 22% of the exam.

13%

Keeping AI systems safe, fair and accurate after launch. About 13% of the exam.

Identifying risks to AI solutions

Monitoring, testing and troubleshooting

Continuous evaluation metrics with Vertex AI Model Monitoring and Explainable AI
Monitoring for training-serving skew
Monitoring for feature attribution drift
Monitoring performance against baselines, simpler models and over time
Monitoring for common training and serving errors