# Vertex AI Platform and Model Garden

> Chapter 11 of the Courseiva GOOGLE-GEN-AI-LEADER curriculum — https://courseiva.com//learn/google-gen-ai-leader/vertex-ai-platform-and-model-garden

**Official objective:** 4.2 — Explain Vertex AI and Model Garden capabilities

## Introduction

Vertex AI Platform and Model Garden. These two services solve a huge problem for anyone trying to use artificial intelligence: they stop you from having to build everything from scratch. If you are studying for the Generative AI Leader exam, understanding these tools is essential because they are Google Clouds main way of making AI practical and accessible for businesses.

## The All-Inclusive Buffet Analogy

20 different kitchen stations line the walls of a massive hall: one for sushi, one for pasta, one for carving meats, one for salads, and so on. Each station has its own trained chef, specialised tools, and ingredients. In a traditional software project, you would have to build and manage each station from scratch — buying the grills, training the chefs, stocking the pantry.

Now imagine a service called Vertex AI Platform. It is the entire buffet hall, already built and staffed. You walk in, pick a station (a tool or service), and start cooking immediately. You do not worry about the kitchen infrastructure, the dishwashers, or the fire safety inspections — that is all included. Model Garden is the menu card at the entrance. It lists every station with a short description: "Station 7: Pre-trained pasta model — just add your sauce." You can select a station, tweak the recipe slightly with your own ingredients (your data), and serve a custom dish without ever entering a kitchen.

But here is the real trick: the buffet also lets you combine stations. You could take a pre-made pizza base from station 3, add special toppings you prepared at station 8, and bake it at station 12. That is what Vertex AI and Model Garden do for artificial intelligence — they give you ready-made AI models and the full environment to customise, combine, and deploy them, without your team having to build a single AI kitchen from zero.

## Core explanation

Vertex AI Platform is Google Clouds unified machine learning (ML) platform. Machine learning is a type of artificial intelligence where computers learn patterns from data instead of being explicitly programmed for every rule. Before Vertex AI, building an ML project meant stitching together many separate Google Cloud services: one for storing data, another for training the model, a third for deploying it, and a fourth for monitoring it. Each service had its own console, its own pricing, and its own way of handling security. This was like trying to build a car by buying the engine from one shop, the wheels from another, and the body from a third, then making them all work together.

Vertex AI Platform bundles all those steps into one place. It covers the entire lifecycle of a machine learning project:

- Data preparation: cleaning and organising raw data so the model can learn from it.
- Training: the process where the model learns patterns from the data. This can take hours or days and requires powerful computers.
- Evaluation: testing the trained model to see how accurate it is before using it with real users.
- Deployment: putting the model online so applications can send it questions and get answers.
- Monitoring: tracking the model's performance over time to catch when it starts making bad predictions.
- Versioning: keeping track of different versions of the model so you can roll back if a new version performs worse.

Model Garden is a feature within Vertex AI Platform. Think of it as a catalogue of pre-built AI models that Google has already trained on enormous amounts of data. These are called foundation models. They are like a student who has already read millions of books — they have a broad general knowledge. Instead of training a new model from scratch (which costs millions of pounds and takes months), you can pick a model from Model Garden and adapt it to your specific need. This process is called fine-tuning. For example, you could take a foundation model that understands English text generally and fine-tune it to answer questions specifically about your companys internal HR policies.

Model Garden includes several types of models:

- Large language models (LLMs): these understand and generate human-like text, such as PaLM and Gemini.
- Vision models: these analyse images and videos, like identifying objects in a photo.
- Chat models: these are optimised for conversational interactions, like chatbots.
- Embedding models: these convert text or images into numerical representations that computers can compare and search through.
- Task-specific models: these are trained for narrow jobs like translation, summarisation, or code generation.

The key insight for the exam is that Vertex AI Platform and Model Garden work together. Vertex AI provides the kitchen (the compute power, storage, and security), while Model Garden provides the recipes (the pre-trained models). Together, they drastically reduce the time, money, and expertise needed to bring AI into a business. A company that previously needed a team of ten data scientists and six months can now achieve the same result with a single AI-literate professional and a few weeks.

Two important concepts for the exam are AutoML and custom training. AutoML lets you train a high-quality model without writing any code — you just upload your data and Vertex AI automatically chooses the best algorithm. Custom training gives you full control, so you can write your own training code using frameworks like TensorFlow or PyTorch. Model Garden supports both approaches: you can use a foundation model as-is (no training needed), fine-tune it with your data (lighter training), or use it as a starting point for custom training.

Finally, Vertex AI Platform handles deployment through what are called endpoints. An endpoint is a web address that your application calls to get predictions from the model. Vertex AI manages the scaling automatically: if one user sends a query, it uses one server; if ten thousand users send queries simultaneously, it spins up hundreds of servers to handle the load. You do not have to worry about servers running out of capacity or sitting idle and wasting money.

## Real-world context

Consider a medium-sized insurance company called SafeGuard Insurance. They want to build a chatbot that can answer customer questions about their policies. The traditional approach would require them to hire ML engineers, collect years of chat transcripts, train a model from scratch, and maintain the infrastructure. Instead, they use Vertex AI Platform and Model Garden.

Step one: the product manager opens Model Garden and browses the available chat models. She selects a foundation model called "chat-bison" which is already good at general conversation. She does not need to understand how it was trained — it just works.

Step two: the team gathers their data — 5,000 past customer service conversations along with the correct answers that agents gave. This data is stored in Google Cloud Storage (a simple file storage service). They upload it to Vertex AI and start a fine-tuning job. Vertex AI automatically provisions the necessary computing power, runs the fine-tuning, and shuts down the resources when done. The team does not need to manage any servers.

Step three: once fine-tuned, the new model is deployed to an endpoint. The endpoint generates a unique URL. The companys website development team configures the live website to send customer chat messages to that URL and display the models responses. Within two weeks, the chatbot is live.

Step four: Vertex AI automatically monitors the model. It tracks metrics like latency (how fast the model responds) and prediction accuracy. If the model starts giving poor answers because the real-world questions differ from the training data, Vertex AI sends an alert. The team can then retrain the model with fresh data.

The result: SafeGuard Insurance saved approximately £200,000 in development costs and launched three months faster than if they had built everything themselves. They did not need to hire specialised AI engineers — their existing IT team could manage the project using Vertex AIs graphical interface.

What does an IT professional actually do with this? They:

- Browse Model Garden to select the right foundation model for the problem.
- Prepare and upload training data in Vertex AI.
- Configure a fine-tuning job (choosing parameters like number of training steps).
- Deploy the model to an endpoint with a single click or API call.
- Set up monitoring dashboards to track model performance.
- Manage access permissions using Cloud IAM to control who in the organisation can modify models.
- Use Vertex AI Pipelines to automate the entire workflow so that retraining happens on a schedule.

## Exam focus

The Generative AI Leader exam tests your understanding of Vertex AI Platform and Model Garden primarily through scenario-based questions. You will not be asked to write code or remember specific API names. Instead, the exam presents a business problem and asks which Google Cloud tool solves it. The traps are designed to catch people who memorise features without understanding their purpose.

Key exam topics to memorise:

- The distinction between a foundation model (pre-trained by Google, used as-is) and a fine-tuned model (adapted with your data).
- The difference between AutoML (no-code training) and custom training (code-based, full control).
- The concept of an endpoint and its role in serving predictions.
- The fact that Model Garden is part of Vertex AI Platform, not a separate product.
- The types of models available in Model Garden: language, vision, chat, embedding, task-specific.
- The term "grounding" — this is a feature where the model checks its responses against a source of truth (like your companys database) to reduce hallucinations (making up false information).

Common exam traps:

- Confusing Vertex AI Platform with Google Clouds general AI offerings like Bard or Gemini. Vertex AI is the platform for building custom AI solutions; Gemini is a specific foundation model accessible through Vertex AI.
- Thinking Model Garden allows you to train models from scratch. It does not — it provides pre-trained models that you can use or fine-tune.
- Assuming that all models in Model Garden are free. They are not; you pay for compute time during training and inference (using the model to make predictions).
- Believing that Vertex AI only works with Googles own models. It also supports open-source models and models you import yourself.

The exam loves to ask: "A company needs to quickly deploy a chatbot that understands their specific product catalogue. Which approach should they use?" The correct answer is always along the lines of: select a foundation chat model from Model Garden, fine-tune it with product data using Vertex AI, and deploy to an endpoint.

Another frequent question type: "What is the primary benefit of Vertex AI Platform over building ML infrastructure from scratch?" The answer revolves around reduced operational overhead, faster time-to-market, and managed scaling — not about it being cheaper (though it often is).

To pass, focus on understanding the "why" behind each feature. For example, why does Vertex AI offer both AutoML and custom training? Because different teams have different needs: a non-technical analyst needs AutoML, while a research scientist needs custom training. The exam wants you to be the leader who can choose the right tool for the right person.

## Step by step

1. **Identify the Business Problem** — Before touching any tool, define what you want the AI to do: answer customer questions, classify images, summarise documents. This step determines whether you need a language model, vision model, or something else, and whether you need to fine-tune or use the model as-is.
2. **Browse Model Garden** — Open Vertex AI Platform in the Google Cloud Console and navigate to Model Garden. Review the available foundation models. For each model, read its description, supported use cases, and pricing. Select the model that best matches your problem domain.
3. **Prepare Your Data** — If you plan to fine-tune the model, gather and clean your training data. This data must be formatted correctly (e.g., as JSON lines or CSV files). Upload the data to Google Cloud Storage, then register it in Vertex AI as a dataset. Vertex AI can help with data labelling if you have unlabelled data.
4. **Configure and Run Training** — In Vertex AI, create a training job. You can use AutoML (no coding needed) for simple cases or custom training if you have specific requirements. Specify the model from Model Garden as the starting point. Set training parameters like the number of epochs (passes through the data). Vertex AI automatically provisions compute resources, runs the training, and shuts down when done.
5. **Evaluate and Deploy the Model** — After training, Vertex AI shows evaluation metrics like accuracy or F1 score. If the performance is acceptable, you deploy the model to an endpoint with one click. Vertex AI creates a secure URL that applications can call. You can also set up A/B testing if you want to compare two model versions before full rollout.
6. **Monitor and Maintain** — Once the model is live, use Vertex AI Model Monitoring to track prediction quality, latency, and data drift (when real-world data starts looking different from training data). Set alerts for when performance drops below a threshold. Periodically retrain the model with fresh data to keep it accurate.

## Comparisons

### Vertex AI Platform vs Model Garden

**Vertex AI Platform:**
- Is the entire environment for ML lifecycle management.
- Includes tools for data prep, training, deployment, and monitoring.
- Can host models you import yourself, not just Google models.

**Model Garden:**
- Is a specific feature within Vertex AI Platform.
- Is only the catalogue of pre-trained foundation models.
- Does not host custom models — their purpose is to provide starting points.

### AutoML vs Custom Training

**AutoML:**
- Requires no coding — works through a graphical interface.
- Google automatically selects the best algorithm for your data.
- Ideal for business analysts and non-ML specialists.

**Custom Training:**
- Requires writing code in Python and ML frameworks.
- Gives full control over the model architecture and training process.
- Ideal for ML engineers and researchers with specific needs.

### Foundation Model vs Fine-tuned Model

**Foundation Model:**
- Trained by Google on massive, general datasets.
- Not optimised for any specific business task.
- Ready to use immediately out of the box.

**Fine-tuned Model:**
- Starts from a foundation model then trained on your data.
- Optimised for your specific business problem (e.g., your product catalogue).
- Requires additional time and compute resources to create.

## Common misconceptions

- **Misconception:** Vertex AI Platform is just a single tool that does one thing. **Reality:** Vertex AI Platform is a unified suite of many tools covering the entire ML lifecycle from data preparation to monitoring. (Because the name 'platform' sounds like a single product, beginners assume it has limited scope, but it actually replaces multiple separate services.)
- **Misconception:** Model Garden is a separate website or service you access independently from Vertex AI. **Reality:** Model Garden is a feature embedded within Vertex AI Platform, accessible through the same console and user interface. (Because 'Garden' sounds like a distinct product name, people think it is a separate portal, similar to how 'App Engine' is different from 'Compute Engine'.)
- **Misconception:** Foundation models in Model Garden are fully trained and ready for any specific business use without any changes. **Reality:** Foundation models are general-purpose and often need fine-tuning with your own data to perform well on your specific task. (Because the word 'foundation' implies completeness, but it actually means a starting point that requires customisation.)
- **Misconception:** You cannot use your own custom models with Vertex AI — you are limited to Googles models in Model Garden. **Reality:** Vertex AI supports importing and deploying your own custom models, including models built with open-source frameworks like TensorFlow and PyTorch. (Because Model Garden is so prominently marketed, beginners assume Vertex AI is a walled garden that only works with Google-owned models.)
- **Misconception:** Fine-tuning a model in Vertex AI requires you to write complex code. **Reality:** Vertex AI offers no-code fine-tuning through its graphical interface, alongside code-based fine-tuning for advanced users. (Because the term 'fine-tuning' sounds technical and programming-heavy, people assume it is inaccessible to non-coders.)

## Key takeaways

- Vertex AI Platform is a single, unified environment that manages the entire machine learning lifecycle from data preparation to model monitoring.
- Model Garden is a catalogue of pre-trained foundation models within Vertex AI, including language, vision, chat, and embedding models.
- Foundation models from Model Garden can be used as-is, fine-tuned with your data, or used as a starting point for custom training.
- Vertex AI endpoints provide a managed URL for your model, with automatic scaling to handle from one to millions of prediction requests.
- AutoML in Vertex AI lets non-coders train high-quality models by simply uploading their data, while custom training gives experts full control.
- The exam focuses on scenario-based questions where you must choose the right Vertex AI tools for a business problem, not on remembering APIs.

## FAQ

**Is Vertex AI the same as Google Cloud AI?**

No. Google Cloud AI is the overall umbrella for all AI services. Vertex AI is one specific product within that umbrella that unifies machine learning tools into one platform.

**Do I need to know how to code to use Vertex AI?**

Not for basic use. Vertex AI has a graphical user interface where you can upload data, select a model from Model Garden, and start training without writing any code. Custom training does require coding, but it is optional.

**What is the difference between a foundation model and a fine-tuned model?**

A foundation model is pre-trained by Google on huge amounts of general data. A fine-tuned model starts from that foundation and is then trained further on your specific data to adapt to your task.

**Can I use my own model with Vertex AI instead of one from Model Garden?**

Yes. Vertex AI supports importing models you built yourself, including models created with open-source frameworks like TensorFlow, PyTorch, or scikit-learn. You are not limited to Model Garden.

**How much does Vertex AI cost?**

Costs vary based on usage: you pay for compute time during training, storage for your data and models, and inference calls when your model serves predictions. Google Cloud publishes a pricing calculator to estimate costs.

**What does 'grounding' mean in Vertex AI?**

Grounding is a feature that makes the models responses more reliable by checking them against a trusted source, like your companys internal database. It reduces the chance of the model inventing false information (hallucinations).

---

Interactive version with quiz and diagrams: https://courseiva.com//learn/google-gen-ai-leader/vertex-ai-platform-and-model-garden