1: Architecting low-code AI solutions

Learning Objectives

By the end of this chapter, you will be able to:

• Choose the correct BigQuery ML model family for a stated business problem across regression, classification, time-series, matrix factorization, boosted trees, and autoencoders, and decide when AutoML Tabular is the better fit.
• Apply feature engineering inside BigQuery ML using the TRANSFORM clause and built-in preprocessing functions, and compare hand-built feature pipelines against AutoML Forecasting.
• Generate predictions with ML.PREDICT, ML.FORECAST, ML.RECOMMEND, and remote-model calls, and debug trained models with ML.EVALUATE, ML.EXPLAIN_PREDICT, and hyperparameter tuning jobs.
• Build applications backed by Model Garden APIs, including Gemini foundation models, Vision, Natural Language, Speech, and Translate.
• Integrate Document AI, Retail API, and Healthcare API into business workloads with appropriate processor and data-store choices.
• Apply prompt engineering patterns (zero-shot, few-shot, chain-of-thought, system instructions, function calling, grounding) to Gemini models on Vertex AI.
• Prepare data for AutoML using Tabular Workflows, data labeling, and feature selection.
• Train custom AutoML models across tabular, text, speech, image, and video modalities.

Executive Summary

• The low-code surface on Google Cloud spans three platforms: BigQuery ML for SQL-native classical ML, AutoML inside Vertex AI for managed multi-modal training, and Model Garden plus industry APIs for pre-trained inference. Choosing among them is the central skill the certification tests.
• BigQuery ML training stays inside the warehouse for any model type listed in its model_type enum, which keeps feature-engineering SQL, training, and prediction in one pipeline. Remote models call out to Vertex AI hosted models or Gemini foundation models per row.
• AutoML compensates for the architectures BigQuery ML does not host, including computer vision, video, custom speech, and high-end tabular models built by Vertex AI Tabular Workflows pipelines under the hood.
• Generative AI on Vertex AI is the path for any prompt-driven workload. The Vertex AI SDK, Generative AI Studio, and Model Garden share the same backend, and the difference between them is interface and billing scope, not capability.

Assumptions

• The reader is a practitioner with at least one year of hands-on Google Cloud experience and is comfortable writing SQL, Python, and basic IAM bindings.
• All region references default to us-central1 unless a service is region-restricted; the discussion of region-restricted APIs (Healthcare API, Document AI processors, Vertex AI generative features) is called out explicitly.
• Service accounts and project structure follow the standard pattern: a workload service account per pipeline with roles/bigquery.user, roles/aiplatform.user, and resource-scoped data-access roles.
• Fictional naming convention throughout the chapter: project acme-ml-prod, dataset acme.retail, model artifacts in acme.models, Vertex AI assets in us-central1.