Skip to content

AI on Google Cloud — Vertex AI, Gemini & Google Cloud AI

What this is: a basic-to-solid tour of Google Cloud's AI portfolio — the unified Vertex AI platform, the Gemini models, and the pre-built Cloud AI services. See the cross-cloud equivalents to map these to AWS/Azure.

1. The Unified Pillar — Vertex AI

Google Cloud's distinctive move: one platform, Vertex AI, covers both generative AI and custom ML (where AWS splits Bedrock/SageMaker and Azure splits Foundry/Azure ML).

Vertex AI does How
Use foundation models (GenAI) Gemini + Model Garden via API and Vertex AI Studio
Build custom ML Training, pipelines, feature store, endpoints — the SageMaker/Azure-ML equivalent

Google's edge: its own Gemini models (top-tier multimodal, very long context), native multimodality, and Grounding with Google Search (no equivalent elsewhere). Strong data-gravity pull if your data already lives in BigQuery.

2. Gemini & the GenAI Side of Vertex

Piece Role
Gemini models Google's flagship LLMs — Gemini 2.5 Pro / Flash — text, image, audio, video in one model; very long context windows
Vertex AI Studio Prompt design, testing, and tuning workbench
Model Garden Catalog of 200+ models — Gemini, plus Llama, Claude (via partners), Mistral, and open models
Vertex AI Agent Builder Build tool-using agents and conversational apps
Vertex AI Search Managed retrieval / RAG over your data (Google-quality search)
RAG Engine / Vector Search Managed RAG orchestration; Vector Search (formerly Matching Engine) for embeddings
Grounding with Google Search Ground answers in live web results to cut hallucination
Gen AI Evaluation Service Pointwise and pairwise judge-based evaluation (details)
Safety filters / Model Armor Configurable content safety and prompt-injection defense

Enterprise properties QA cares about

IAM, VPC Service Controls, data residency, CMEK, and no-training-on-your-data terms.

3. The Custom-ML Side of Vertex AI

The classic-ML platform (equivalent to SageMaker / Azure ML), all under the same Vertex roof:

Piece Role
Workbench Managed Jupyter notebooks
Training Custom + AutoML training on managed compute
Pipelines MLOps orchestration (Kubeflow-based)
Feature Store Central feature management/serving
Model Registry & Endpoints Versioning + real-time/batch deployment
Model Monitoring Drift and skew detection in production
Vertex Explainable AI Feature attributions / explainability

4. Pre-Built Cloud AI Services (No Training Needed)

Service Does
Document AI OCR / structured document extraction
Vision AI Image analysis, OCR, object/label detection
Speech-to-Text / Text-to-Speech Speech ↔ text
Cloud Translation Machine translation
Cloud Natural Language AI NLP — entities, sentiment, syntax
Vertex AI Search Enterprise semantic search
BigQuery ML Train & run ML directly in SQL inside BigQuery — a GCP signature capability

Assistant: Gemini for Google Workspace (productivity) and Gemini Code Assist (developers).

5. A Typical GCP GenAI Architecture

flowchart LR
    U["App / user"] --> IAM["Cloud IAM auth"]
    IAM --> GEM["Vertex AI / Gemini"]
    DOCS["GCS / BigQuery data"] --> VS["Vertex AI Search /<br/>Vector Search (embed + index)"]
    GEM <--> VS
    GEM --> GRND["Grounding w/ Google Search"]
    GEM --> EVAL["Gen AI Evaluation (CI gate)"]

6. QA / Testing Focus on GCP

  • Gemini calls — quality (Gen AI Eval, pairwise for prompt A/B), latency, token cost.
  • Vertex AI Search / RAG Engine — retrieval quality, grounding accuracy, RAG metrics.
  • Agents (Agent Builder) — tool-call correctness, task completion (AI Test Strategy).
  • Grounding — verify web-grounded claims actually cite retrieved sources (anti-hallucination).
  • IAM & residency — VPC Service Controls, region-boundary tests (AI Rollout Part 1).

Where to Go Next