Learning Objectives

• Design and deploy multi-turn conversational and agentic applications using Vertex AI Agent Builder and the Gemini API
• Implement retrieval-augmented generation pipelines using Vertex AI Search and vector-based data stores
• Fine-tune Gemini models on domain-specific datasets using supervised fine-tuning workflows on Vertex AI
• Apply Vertex AI Model Evaluation and the Gen AI Evaluation SDK to measure model quality across custom metrics
• Integrate responsible AI tooling including safety filters, grounding checks, and model monitoring into production systems
• Orchestrate multi-agent workflows using LangChain, LangGraph, or native Vertex AI extensions and tool use
• Build and manage Vertex AI Pipelines for repeatable training, evaluation, and deployment workflows
• Configure CI/CD patterns for generative AI applications using Cloud Build and Artifact Registry
• Apply cost, latency, and throughput optimization strategies when serving Gemini models at scale
• Architect end-to-end solutions that meet enterprise security, compliance, and observability requirements on Google Cloud

Course Outline

Module 1: Gemini API Deep Dive

Reviews the Gemini model family (Gemini 1.5 Pro, Flash, Nano) and their capability differences. Covers the Vertex AI SDK and REST API in depth, including multimodal inputs (text, image, video, audio, documents), system instructions, and advanced generation parameters. Students work through token budgeting, context window management, and streaming responses in Python.

Module 2: Advanced Prompt Engineering and Structured Output

Covers chain-of-thought, few-shot, and role-based prompting at an advanced level. Introduces controlled generation using JSON mode and response schemas to produce structured outputs reliably. Discusses prompt versioning strategies and how to manage prompt templates across environments using Vertex AI Prompt Management.

Module 3: Function Calling and Tool Use

Examines Gemini’s function calling capability in detail, including defining tool schemas, handling parallel function calls, and managing multi-turn tool-use conversations. Students build a working example that connects Gemini to external APIs and a mock database, then explore error handling and fallback patterns when tool calls fail or return unexpected results.

Module 4: Vertex AI Agent Builder — Core Concepts

Introduces Vertex AI Agent Builder, including the Dialogflow CX integration, data store agents, and playbooks. Covers how to create grounded agents that retrieve information from connected data sources. Students configure a basic customer-facing agent, wire it to a data store, and test grounding behavior, citations, and fallback responses.

Module 5: Multi-Agent Architectures and Orchestration

Covers patterns for composing multiple specialized agents into a coordinated system, including orchestrator-worker patterns and agent handoffs. Introduces LangGraph for stateful multi-agent workflows and compares it with native Vertex AI Extensions. Students design and implement a multi-agent pipeline where a router agent delegates subtasks to domain-specific subagents.

Module 6: Retrieval-Augmented Generation with Vertex AI Search

Covers the end-to-end architecture of RAG on Google Cloud using Vertex AI Search as the retrieval layer. Topics include creating and populating data stores (website, Cloud Storage, BigQuery, and structured data sources), configuring chunk size and embedding models, and integrating retrieval results into Gemini generation calls. Students evaluate retrieval quality and tune ranking and filtering settings.

Module 7: Vector Search and Custom Embedding Pipelines

Introduces Vertex AI Vector Search (formerly Matching Engine) for low-latency approximate nearest neighbor retrieval. Covers generating embeddings with Vertex AI text-embedding models, indexing documents at scale, and querying the index at runtime. Students build a custom RAG pipeline using Vector Search as an alternative to Vertex AI Search, comparing the trade-offs between the two approaches.

Module 8: Supervised Fine-Tuning on Vertex AI

Covers when fine-tuning is justified versus prompt engineering or RAG, and how to prepare high-quality training datasets in JSONL format. Students walk through the full supervised fine-tuning workflow for Gemini on Vertex AI: dataset upload to Cloud Storage, tuning job configuration, monitoring training loss, and deploying the tuned model endpoint. Discusses data volume requirements and common failure modes.

Module 9: Model Evaluation and the Gen AI Evaluation SDK

Introduces Vertex AI Model Evaluation and the Gen AI Evaluation SDK for assessing generative model outputs. Covers built-in metrics (fluency, groundedness, safety, summarization quality) and how to define custom pointwise and pairwise evaluation metrics using a judge model. Students run evaluation pipelines comparing a base Gemini model against a fine-tuned variant and interpret the results.

Module 10: Responsible AI Tooling and Safety Controls

Examines Vertex AI’s responsible AI features in depth, including configurable safety filters, harm categories, and threshold tuning. Covers grounding with Google Search to reduce hallucination, and the use of model cards and data cards for documentation. Students integrate safety checks into an application pipeline and review how to interpret and act on safety attribute scores returned by the API.

Module 11: Vertex AI Pipelines for Generative AI Workflows

Introduces Kubeflow Pipelines on Vertex AI as a tool for building repeatable, auditable ML workflows. Students author pipeline components for data ingestion, embedding generation, fine-tuning job submission, evaluation, and conditional deployment. Covers pipeline scheduling, caching, and artifact lineage tracking in Vertex ML Metadata.

Module 12: Serving, Scaling, and Cost Optimization

Covers deployment options for Gemini-based applications, including Vertex AI Endpoints, Cloud Run, and API Gateway. Discusses autoscaling configuration, request batching, quota management, and caching strategies to reduce latency and cost. Students profile a sample application to identify bottlenecks and apply optimizations, including model selection trade-offs between Gemini Pro and Flash for cost-sensitive workloads.

Module 13: MLOps and CI/CD for Generative AI

Covers the principles of MLOps applied specifically to generative AI: prompt versioning, model registry management, A/B testing endpoints, and automated evaluation gates in CI/CD pipelines. Students configure a Cloud Build pipeline that runs evaluation checks before promoting a new model or prompt version to production, using Artifact Registry to version pipeline components and container images.

Module 14: Observability, Monitoring, and Drift Detection

Examines how to monitor production Gemini applications using Vertex AI Model Monitoring, Cloud Logging, and Cloud Monitoring. Covers logging prompts and responses for audit purposes (with privacy considerations), setting up alerts on latency and error rate metrics, and detecting distributional drift in inputs over time. Students configure a monitoring dashboard and define alert policies for a deployed agent.

Module 15: Security, IAM, and Enterprise Governance

Covers security controls relevant to production Vertex AI deployments: VPC Service Controls, private endpoints, customer-managed encryption keys (CMEK), and organization-level AI usage policies. Reviews IAM roles specific to Vertex AI and Agent Builder, and discusses data residency, audit logging with Cloud Audit Logs, and compliance considerations for regulated industries.

Module 16: Capstone Project — End-to-End Production System

Students design and build a complete production-grade application incorporating the major components covered throughout the course: a grounded multi-agent system with RAG, a fine-tuned or specialized model endpoint, an evaluation pipeline, safety controls, and a CI/CD deployment workflow. Each project is reviewed against a rubric covering architecture quality, responsible AI compliance, observability, and documentation.