13.2 Prebuilt AI APIs, Model Garden, Gemini Cloud Assist & Securing AI

Prebuilt AI APIs, Model Garden, Gemini Cloud Assist & Securing AI

Architectural Objective: Blueprint 2.5 (Configuring prebuilt solutions or APIs with Agent Platform) expects you to differentiate Google's AI APIs, integrate Gemini Enterprise features, and plug Model Garden models into solutions. Blueprint 3.1 adds Securing AI — Model Armor, Sensitive Data Protection, and secure model deployment — and Gemini Cloud Assist appears in Domains 1 and 5 as the design and operations assistant.

---

Differentiating the Prebuilt Google AI APIs

The blueprint enumerates API categories — Search, Conversation, Vision, Image, Video, and Audio. The exam asks you to pick the right one from scenario cues:

Model Garden: Choosing the Foundation Model

Model Garden (within Vertex AI) is the curated catalog of deployable foundation models:

• First-party: the Gemini family for multimodal reasoning and generation; Gemma open models for lightweight, portable workloads; task families such as Imagen (image).
• Third-party and open: select partner and open-weight models available for deployment on Vertex AI endpoints with Google Cloud security and IAM posture.
• Exam lens: Model Garden is the answer when a scenario needs to select, evaluate, and deploy an existing foundation model quickly rather than train from scratch — combined with tuning (supervised fine-tuning, RAG grounding) to specialize it.

Gemini Enterprise Features in Enterprise Workflows

Blueprint 2.5 names AI Agents and NotebookLM as Gemini Enterprise features to integrate:

• AI agents: task-completing assistants grounded in enterprise data — customer service, research, and workflow automation agents composed and governed in the Gemini Enterprise Agent Platform.
• NotebookLM: the AI research and synthesis assistant that grounds answers in a user's own source documents, useful for knowledge-work acceleration without building anything.

Gemini Cloud Assist: AI for the Infrastructure Itself

Gemini Cloud Assist is Google Cloud's AI assistant for the cloud platform — appearing in blueprint 1.2 (designing infrastructure that meets technical requirements) and 5.1 (advising teams):

• Design assistance: architecture and configuration recommendations aligned to best practices.
• Operations: troubleshooting help interpreting errors, logs, and performance signals; cost and optimization insights.
• Exam lens: when a scenario asks how architects or operators can investigate a production issue or optimize a design faster with built-in AI assistance in Google Cloud, Gemini Cloud Assist is the intended answer — distinct from Vertex AI, which builds custom models.

---

Securing AI (Blueprint 3.1)

Generative AI opens a new attack surface; the blueprint's security domain now explicitly includes securing AI:

+-----------------------------------------------------------------------------------+
|                          SECURING-AI CONTROL STACK                                |
+-----------------------------------------------------------------------------------+
| PROMPT LAYER   | Model Armor: screens prompts/responses for prompt injection,     |
|                | jailbreaks, malicious URLs, and harmful content.                 |
| DATA LAYER     | Sensitive Data Protection: discovers, classifies, and de-        |
|                | identifies sensitive data in prompts, groundings, training sets. |
| DEPLOYMENT     | Private endpoints (Private Service Connect), IAM least privilege,|
|                | VPC Service Controls perimeters around models and training data. |
| GOVERNANCE     | Model Registry versioning, evaluation gates, Explainable AI,     |
|                | audit logging of model access.                                   |
+-----------------------------------------------------------------------------------+

• Model Armor is the dedicated runtime defense for generative AI: prompt-injection and jailbreak detection, harmful-content filtering, sensitive-data detection and grounding checks — position it in front of LLM-facing endpoints.
• Sensitive Data Protection (SDP) scans the data flowing into prompts, fine-tuning sets, and agent responses so PII/PHI is discovered, classified, and masked.
• Secure model deployment follows standard architecture discipline: private connectivity (Private Service Connect) instead of public endpoints, least-privilege service accounts, VPC Service Controls perimeters to prevent training-data exfiltration, and CMEK where the enterprise mandates key custody.

---

Concrete Scenario: Insurer Claims-AI Platform

An insurer wants a claims agent (Conversation), damage-photo assessment (Vision), invoice parsing (Document AI), and a proprietary fraud model (custom Vertex AI with explainability):

1. APIs first: Conversational Agents + Vision AI + Document AI cover three of the four needs with zero training.
2. Custom fraud model: trained in Vertex AI from historical claims, deployed behind a private endpoint with Explainable AI for regulator-facing reason codes.
3. Security: Model Armor in front of the claims agent; SDP masks policyholder data in prompts and logs; VPC-SC perimeter around the training project.

---

[!IMPORTANT] Exam Watch: Match the cue to the API family (search, conversation, vision, image, video, audio) before reaching for custom training; choose Model Garden to deploy an existing foundation model; name NotebookLM and AI agents when the requirement is workforce productivity rather than product features; answer infrastructure-assistance questions with Gemini Cloud Assist; and for any generative AI endpoint exposed to users, include Model Armor + Sensitive Data Protection + private/IAM-secured deployment.