Cloud Pro: Google Cloud AI and Machine Learning: Choose, Compare and Launch

What's inside

1. Level 1: The Google Cloud AI Map, Old Names and NewFree
   Every Google Cloud AI and ML service sorted onto one page by group: generative platforms and models, agents, search and retrieval, ready-made AI APIs, ML platforms and accelerators. Each one shows its current name and its former names, including Vertex AI becoming Gemini Enterprise Agent Platform at Next '26 in April 2026 and Agentspace becoming Gemini Enterprise.
2. Level 2: The Standard AI Workload ProfileFree
   A reusable profile that describes any AI workload before you pick a service: the task type, whether to buy, adapt or train, data sensitivity and residency, latency and throughput, volume and token or request cost, lock-in, evaluation and safety, MLOps maturity, accelerators, team skills and cost model.
3. Level 3: The Launch Kit: Project, IAM, Quotas, gcloud, Terraform
   The shared base that every quick start in the course reuses: a sandbox project with billing, the APIs you need, least-privilege service accounts, quota checks, gcloud and Terraform configuration, budget alerts and a dependable clean-up routine.
4. Level 4: Model Garden: First-Party, Partner and Open Models
   The nine-part lesson for Model Garden: how Google, partner and open models are offered (managed API versus self-deployed endpoint), the licence and regional limits that bite, and who runs what.
5. Level 5: Gemini Models: Family, Token Limits and Versions
   The Gemini model family on Google Cloud: tiers, context windows, input and output token limits, multimodal inputs, regional and global endpoints, and the version lifecycle that retires model IDs on a published schedule.
6. Level 6: Gemini Pricing, Caching and Provisioned Throughput
   Where generative AI spend comes from and how to control it: input and output token prices, long-context tiers, context caching, batch inference, pay-as-you-go quotas compared with provisioned throughput, and the 429 errors that push teams to reserve capacity.
7. Level 7: Tuning Gemini and Open Models
   When and how to adapt a model rather than just prompt it: supervised fine-tuning for Gemini, parameter-efficient tuning for open models, preparing a dataset, what tuning jobs and tuned endpoints cost, and the cases where tuning loses to better prompts or retrieval.
8. Level 8: Agent Builder and the Agent Development Kit
   Building agents in code with the Agent Development Kit (ADK) inside Agent Builder: models, tools, function calling, multi-agent setups and agent-to-agent (A2A) interop, plus the design mistakes that make agents slow, expensive or unsafe.
9. Level 9: Agent Engine: Deploying and Running Agents
   The managed runtime for agents: deploying an ADK agent, sessions and memory, scaling, identity and observability, how runtime is billed, and what you still operate yourself compared with running the agent on Cloud Run.
10. Level 10: Gemini Enterprise in Brief
    Gemini Enterprise (formerly Agentspace) as the bought, per-seat workplace product for employee search, assistants and agents across company apps, and how it differs from building on Agent Platform.
11. Level 11: Vertex AI Search (Agent Search)
    The nine-part lesson for managed enterprise search: data stores, website, unstructured and structured sources, search apps, grounded answers, query and indexing pricing, and limits that bite on document size and freshness.
12. Level 12: Conversational Agents (Dialogflow CX)
    The nine-part lesson for Conversational Agents: deterministic flows and pages compared with generative playbooks, data store answers, telephony and contact centre integration, per-request and per-session pricing, and the limits that bite.
13. Level 13: Decision Drill: Generative, Agent or Search Service
    A timed drill that runs profiled workloads through every service from this phase (Model Garden, Gemini, tuning, ADK, Agent Engine, Gemini Enterprise, Search, Conversational Agents) and justifies one choice for each.
14. Level 14: Custom Training Jobs
    Running your own training code on managed infrastructure: prebuilt and custom containers, machine and GPU choices, distributed training, Spot VMs, hyperparameter tuning, and what billing per node-hour means for your budget.
15. Level 15: Online Prediction Endpoints
    Serving models in real time: Model Registry, endpoints, machine types and GPUs, autoscaling and minimum replicas, traffic splitting, and the always-on node-hour cost that catches out new teams.
16. Level 16: Batch Prediction and Serving Cost Control
    Offline scoring with batch prediction jobs from Cloud Storage or BigQuery, and how to choose between batch, online endpoints and serverless options based on volume and latency.
17. Level 17: Pipelines: Repeatable ML Workflows
    Managed pipelines with the Kubeflow Pipelines SDK: components, caching, artifacts and lineage, scheduled runs, what each run costs, and the line between a pipeline and a simple scheduled job.
18. Level 18: Feature Store V2 and Leaving the Legacy Version
    Feature Store V2 built on BigQuery: feature groups, online stores (Bigtable or optimized serving), feature views and sync, plus the migration path off Feature Store (Legacy), which was deprecated on 17 February 2026 and shuts down on 17 February 2027.
19. Level 19: BigQuery ML
    Training and running models in SQL inside BigQuery: built-in model types, imported and remote models including Gemini through ML functions, how on-demand and capacity pricing apply to ML queries, and where SQL-first ML reaches its limits.
20. Level 20: Cloud TPU in Brief
    Cloud TPUs compared with GPUs for training and serving: generations and slices, on-demand, Spot and reserved capacity, framework fit (JAX, PyTorch/XLA), availability limits, and when accelerators are worth the engineering effort.
21. Level 21: Document AI: Processors, Pricing and Gemini Rivals
    Learn to choose and launch a Document AI processor (OCR, Form Parser, Layout Parser, specialised and custom extractors) and to judge it against Gemini reading the PDF directly, using page-based pricing, page and file-size limits, and the effort each needs to operate.
22. Level 22: Speech-to-Text: V2, Chirp Models and Recognizers
    Learn to transcribe audio with Speech-to-Text V2 by picking a model (including Chirp 3), recognition mode (synchronous, streaming or batch), region and recognizer, and to weigh per-minute costs and audio-length limits against Gemini audio understanding.
23. Level 23: Text-to-Speech: Voice Tiers and Gemini-TTS
    Learn to pick a Text-to-Speech voice tier (Standard, WaveNet, Neural2, Chirp 3: HD, Gemini-TTS) by quality, controllability and per-character price, and to make a working synthesis call within the request-size and SSML limits.
24. Level 24: Vision API: Feature Pricing and Image Limits
    Learn to call the Cloud Vision API for labels, OCR, SafeSearch and object detection, price it per feature per image, live within its image-size and batch limits, and know when Gemini's multimodal input or a custom-trained model should replace it.
25. Level 25: Translation: NMT, Translation LLM and Glossaries
    Learn to use Cloud Translation Advanced by choosing among the standard NMT model, the Translation LLM and adaptive translation, controlling terminology with glossaries, and pricing per character against general Gemini translation.
26. Level 26: Natural Language API and When Gemini Wins
    Learn what the Cloud Natural Language API still does well (entities, sentiment, classification, moderation) at its per-unit pricing, which related services are deprecated, and how to decide when a Gemini prompt beats any of the dedicated ready-made AI APIs, and when it loses.
27. Level 27: Choosing a Model: Size, Family and Openness
    Learn a repeatable way to narrow the model shortlist for a task by trading Gemini tiers, partner models and open-weight models from Model Garden against one another on capability, context window, latency, cost per token, regions and lock-in.
28. Level 28: Evaluating Models with Your Own Test Set
    Learn to evaluate shortlisted models against a golden dataset using the platform's generative AI evaluation service, computed metrics and model-based judges, so model choice and prompt changes rest on evidence rather than on demos.
29. Level 29: RAG Architecture: Build, Buy or Ground
    Learn the main retrieval-augmented generation designs on Google Cloud (managed Agent Search / Vertex AI Search, the RAG Engine, a do-it-yourself vector store in Vector Search, AlloyDB or BigQuery, and Google Search grounding) and when each fits.
30. Level 30: RAG Quality: Chunking, Ranking and Citations
    Learn the levers that decide RAG answer quality in production (parsing, chunk size, embeddings, hybrid search, reranking, metadata filters and citations) and how to measure retrieval and groundedness separately.
31. Level 31: Responsible AI and Safety Filters
    Learn to configure Gemini safety settings and thresholds, add Model Armor-style prompt and response screening, defend against prompt injection in agents, and document responsible-AI decisions so they stand up to review.
32. Level 32: Data Protection, Residency and Governance
    Learn how customer data is handled by the AI services (training-use commitments, caching, abuse logging), and how to meet residency and security needs with regional endpoints, CMEK, VPC Service Controls, IAM and audit logs.
33. Level 33: MLOps and LLMOps Maturity
    Learn to assess and raise MLOps maturity across predictive and generative workloads (versioned data, prompts and models, pipelines, model registry, monitoring for drift and quality, and a controlled rollout and rollback).
34. Level 34: Worked Cost: One Generative, One Predictive
    Learn to build two complete dated cost models, a Gemini-based RAG assistant (tokens, caching, grounding, pay-as-you-go vs provisioned throughput) and a predictive model (training, endpoint nodes vs batch, BigQuery ML), and find the biggest cost levers.
35. Level 35: Building the Weighted Selection Scorecard
    Learn to turn the standard AI workload profile into a weighted scorecard with must-pass gates, weighted criteria and explicit scoring rules, so a service choice can be defended and repeated.
36. Level 36: Cross-Cloud Evidence: Mapping Equivalent Services
    Learn to map each Google Cloud AI service to its nearest equivalent on another cloud (for example AWS Bedrock and SageMaker, or Azure AI Foundry) and to collect dated, sourced evidence on models, limits, prices and regions without comparing unlike things.
37. Level 37: Scoring Google Cloud Against Another Cloud
    Learn to score Google Cloud and a second cloud against the same scorecard using the dated evidence, including exit and lock-in costs, team skills and existing commitments, and show how robust the result is.
38. Level 38: Final Challenge: Decision Workshop and Record Exam
    Apply the whole course to a realistic brief: profile the workload, shortlist and evaluate services, cost them, check safety and residency, score Google Cloud against another cloud, plan the launch and clean-up, and write a one-page selection record.

Access

The first 2 levels are free with a free account. Every level, the podcast edition and the AI tutor come with All Access at £4.99/month or any Creator plan — see pricing.

Start free →