DAY 04 · MODULE 3

Four ways to build. 🛤️

Back to predictive AI. Four ways to build any ML model — from no-code to do-it-yourself — and how to pick the right one.

Module 3 — 5 Chapters 🗺️

The official AI Development Options course. Compare the four ways to build, then go deep on each.

🧭

1 · The 4 Options

Compare & choose

🏭

2 · Vertex AI

The unified platform

🤖

3 · AutoML

No-code ML

🔌

4 · Pre-trained APIs

Borrow Google's models

🧪

5 · Custom Training

Full control · TensorFlow

One big idea runs through it — match the tool to your skills, your data, and your budget. Jom.

Chapter 1 · The 4 Options

Four Ways to Build an ML Model 🧭

Module 2 was generative AI. Now back to predictive AI — forecasting & classification. Four paths, easiest to hardest:

🔌

Pre-trained APIs

Call Google's ready models. No training data. Lowest effort.

📊

BigQuery ML

Build models with SQL, where your data lives.

🤖

AutoML

Bring your data, point & click. No model code.

🧪

Custom Training

Write the code. Full control. Highest effort.

The golden rule (again)Meet yourself where you are. Climb only as high as your skills & needs require — no higher.

Compare Them Side by Side 📊

The exam loves this comparison. Five things that separate the four options.

Option	Data types	Training data	Expertise	Tune knobs?	Train time
🔌 Pre-trained API	All + audio	None	Lowest	No	None
📊 BigQuery ML	Tabular + JSON	Large	SQL	Yes	Varies
🤖 AutoML	Tabular + image	Your own	Low	No	Varies
🧪 Custom Training	Tabular/img/text/video	Large	Highest	Yes	Longest

Quick decision guide: no data / little ML → Pre-trained API · SQL & data in BigQuery → BigQuery ML · custom model, minimal code → AutoML · full control → Custom training.

Try It · Chapter 1

Which Option Should You Use? 🌳

🕹️Tap an answer to each question — the tree branches to a recommendation. Goal: see how data, skills & control decide your tool. Reach an answer → +5⭐ (try other paths too).

No single "best" — it depends on your data, your skills, and your budget.

Chapter 2 · Vertex AI

One Unified Platform 🏭

Building ML is hard — per Gartner, only half of enterprise ML projects ever get past the pilot. Vertex AI fixes that.

What "unified" meansOne platform for the whole ML workflow — and it covers both generative AI and predictive AI.

🤖

AutoML

No-code path inside Vertex AI

🧪

Custom Training

Code-based path inside Vertex AI

🔗

Workbench

Even write SQL to connect BigQuery + Vertex AI

The End-to-End ML Pipeline 🔄

Vertex AI walks you from raw data to a monitored model in production:

1

Data Readiness

Upload from Cloud Storage, BigQuery, or local

→

2

Feature Readiness

Create features · share via Feature Store

3

Train & Tune

Experiment with models & hyperparameters

→

4

Deploy & Monitor

Into production · auto-monitor & improve (MLOps)

Same four beats you'll see again in Module 4's workflow. Data → features → train → deploy.

Why Vertex AI? The 4 S's ✨

🪄

Seamless

Smooth from uploading data to model in production.

📈

Scalable

MLOps auto-scales storage & compute.

♻️

Sustainable

Artifacts & features can be reused & shared.

⚡

Speedy

~80% fewer lines of code than competitors.

RememberSeamless · Scalable · Sustainable · Speedy. Four S's.

Chapter 3 · AutoML

AutoML — No-Code ML 🤖

AutoML = Automated Machine Learning. It does the time-consuming parts for you.

The promiseBring your data, point & click. AutoML handles feature engineering, model search, tuning & ensembling — so you focus on the business problem, not the model architecture.

Born in 2018 to save data scientists' manual work; since 2021 it's built into Vertex AI.

How AutoML Works — 4 Phases ⚙️

1

Data Processing

Auto-converts numbers, dates, text, categories

→

2

Search & Tune

Finds best models, tunes parameters

3

Ensemble

Assembles the top ~10 models (not just one)

→

4

Prediction

Serves the combined result

Why ensemble? Averaging the top ~10 models beats relying on a single one — better accuracy.

The Two Engines Behind It — Tap 🔍

🕹️Tap each card to open it. Goal: see the two techniques that power AutoML — Neural Architecture Search & Transfer Learning. (Explore — no points.)

🏗️ Neural Architecture Search

tap to expand

Tries many different model architectures, compares their performance, and automatically picks the best for your data — tuning parameters as it goes.

♻️ Transfer Learning

tap to expand

Reuses pre-trained models as a starting point — like learning new things from existing knowledge. Reaches higher accuracy with less data & compute. (Same idea as LLMs: pre-train, then fine-tune.)

Together they let AutoML find strong models fast — even with smaller datasets.

Chapter 4 · Pre-trained APIs

Borrow Google's Models 🔌

Good models need lots of data — hundreds of thousands of records. No data? Pre-trained APIs are the place to start.

API = a power outlet 🔌You just need the right adapter — which API to call and what to pass. You don't worry about the wiring behind the wall (the model training & deployment).

API = Application Programming Interface — how software talks to software. Call a pre-trained model like calling a function. No training data, no ML expertise needed.

Try It · Chapter 4

Call the Gemini API — Tap Each Step 🧪

🕹️Tap each highlighted line of Python → what that step does. Goal: see that calling a model is just 4 steps — authenticate, pick, send, read. Each step → +2⭐.

genai.configure(api_key="YOUR_API_KEY") model = genai.GenerativeModel("gemini-2.5-flash") response = model.generate_content("Top 3 largest countries by area?") print(response.text)

Tap a highlighted line to see what that step does.

No model training. You call a ready model like a function. Four steps — authenticate, pick model, send prompt, read reply.

The API Menu 📋

✨

Gen AI APIs

Foundation models — the multimodal Gemini APIs to create content.

🏭

ML APIs

e.g. the Vertex AI API — train, monitor & tune with minimal effort.

🗣️

Perception APIs

Speech, image, document & conversation APIs.

The Natural Language API analyses text right in your browser — identify entities, sentiment, syntax & categories. That's today's hands-on lab. Many of these are increasingly replaced by the multimodal, multi-task Gemini APIs.

Chapter 5 · Custom Training

Custom Training — Full Control 🧪

When your needs go beyond AutoML's automation, you do it yourself — full control of architecture, frameworks & training logic.

First decision: your containerBefore any code, choose the environment your training runs in — a pre-built container, or a custom one.

Pre-built vs Custom Container 🍳

🍳 Pre-built Container

a furnished kitchen

Cabinets, appliances, cookware ready
Python, TensorFlow, PyTorch already set up
Best when you're not fussy about the infra

🏚️ Custom Container

an empty room

You define every appliance & tool
Set the environment, machine type, disks
Best when you need exact control

Where you code: Vertex AI Workbench (managed Jupyter for the whole workflow) or Colab Enterprise (integrated into Vertex AI in 2023).

Libraries & TensorFlow 🧰

You don't code from scratch — you use ML libraries. Popular, open-source ones:

🔶

TensorFlow

Google's end-to-end platform

🟠

scikit-learn

Classic ML

🔥

PyTorch

Popular for research

⚡

JAX

High-performance, flexible

TensorFlow's layers (low → high)

Keras — high-level API hides the details · most-used
Model libraries neural-net layers, metrics
Low-level APIs + Hardware C++/Python · CPU · GPU · TPU

tf.keras in 3 Steps 🪜

Building a model with the high-level Keras API always follows the same three beats:

1

Create

Piece together the neural-net layers · tf.keras.Sequential

→

2

Compile

Set hyperparameters — loss function, optimizer

→

3

Fit (Train)

Train over epochs to find the best fit

Create, compile, fit. Then deploy & predict. Vertex AI fully hosts TensorFlow, so it's a managed service whichever layer you code at.

Knowledge Check · Module 3

Lock It In 🧪

Q1Which option needs NO training data?

Pre-trained APIs

BigQuery ML

AutoML

Custom training

Q2Want a custom model but minimal coding. Pick:

Pre-trained APIs

Custom training

AutoML

BigQuery ML

Q3The first stage of the Vertex AI pipeline is:

Data readiness

Deployment

Monitoring

Prediction

Q4AutoML's standout feature is:

It needs no data

A no-code, point-and-click UI

It's a chip

It only writes SQL

Q5Since 2021, AutoML lives:

Inside Vertex AI

Inside BigQuery only

On your laptop only

In Google Sheets

Q6The first step when calling the Gemini API is to:

Train a model

Print the response

Authenticate with your API key

Upload a dataset

Q7The most-used, high-level TensorFlow API is:

C++

SQL

Keras

JAX

Q8Choose custom training when you need:

Full control of architecture & training logic

No code at all

No data at all

Only a SQL query

Module 3 — You Can Now Explain… ✅

Pre-trained API BigQuery ML AutoML Custom Training

Vertex AI (unified) Data→Feature→Train→Deploy 4 S's

Neural Architecture Search Transfer Learning Ensemble

API = adapter Pre-built vs Custom Container TensorFlow · Keras Create·Compile·Fit

Finish Module 3 in 2 stepsDo the hands-on lab first — then take the graded quiz.

🧪 Step 1 · Lab — Entity & Sentiment Analysis with the Natural Language API

Call a pre-trained API to pull entities & sentiment from real text. Open the lab →

📝 Step 2 · Official Module 3 Quiz →

After the lab, take the graded quiz on Skills Boost. skills.google › course 593 › quiz 617905

Four ways to build — and you can pick. 🛤️

MODULE 3 COMPLETE

Next: AI Development Workflow. 🔄

You know the four ways to build. Last module — the end-to-end workflow: data prep, how a model learns, training, evaluation, deployment, and MLOps.

🔄

M4 · AI Dev Workflow

Start now →

🏅

Then: the exam

Almost there

🎉

Day 4 done

So close