DAY 04 · AI & ML ON GOOGLE CLOUD

Today is easier than yesterday. 🛠️

Five quick modules. The hands-on side of AI on Google Cloud. By 5pm you'll know what to click and why.

Here's What We'll Cover 🗺️

🥞

M1 · The Stack

Google's 4 AI layers

🛤️

M2 · Pick a Path

API vs AutoML vs BQ ML vs Custom

🪄

M3 · BigQuery ML

ML with pure SQL

🏭

M4 · Vertex AI

One platform, all your ML

🔄

M5 · MLOps

Keep models healthy in production

🏅

Then: Homework

4 labs on Skills Boost

Same deal as Day 3. Framework + two labs in class, rest as homework. You'll be ready.

Before We Dive In

Today's Labs — 2 in Class, 2 at Home 🧪

All labs live on Skills Boost. Open it now in another tab: skills.google → sign in with Google → find "Introduction to AI and Machine Learning on Google Cloud".

🎯

Live in class

Entity & Sentiment Analysis with NL API (~45 min)
Get Started with Agent Studio (~60 min)

🏠

Homework tonight

Predict Visitor Purchases with BigQuery ML (~60 min)
Predicting Loan Risk with AutoML (~60 min)

Why split it4 hours of clicking in one day fries the brain. The 2 we do together are the most fun ones. The 2 at home are more SQL-heavy and easier to do at your own pace.

All 4 labs count toward the official Google badge. The split is for sanity, not skipping.

MODULE 01

The 4-Layer Stack 🥞

Four layers. Remember them in order. That's the whole module.

Just 4 Layers 🥞

Layer 4 — AI Solutions Contact Center AI · Document AI

Layer 3 — Pre-trained APIs Vision · Speech · Translate · NL

Layer 2 — ML Development Vertex AI · BigQuery ML · AutoML

Layer 1 — Infrastructure TPUs · GPUs · Storage

Golden RuleClimb only as high as you need. Pre-trained APIs first. Custom training last.

Try It · Module 1

Match Each Product to Its Layer 🥞

Drag each Google product onto the layer where it lives.

TPU v5p chips

BigQuery ML

Vision API

Contact Center AI

Vertex AI (Agent Platform)

Speech-to-Text

Document AI

GPUs

4️⃣ AI Solutions

3️⃣ Pre-trained APIs

2️⃣ ML Development

1️⃣ Infrastructure

Pattern: chips = bottom · Vertex/BQ ML = middle · APIs = pre-trained · "X AI" solutions = top.

Knowledge Check · Module 1

5 Quick Wins 🧪

Q1Vertex AI sits in which layer?

Infrastructure

ML Development

Pre-trained APIs

AI Solutions

Q2Best layer to start with for a generic problem?

Custom training

Pre-trained APIs

Infrastructure

AutoML

Q3Contact Center AI lives in which layer?

Infrastructure

ML Development

Pre-trained APIs

AI Solutions

Q4TPUs and GPUs belong to which layer?

Infrastructure

ML Development

Pre-trained APIs

AI Solutions

Q5AutoML and BigQuery ML both live in which layer?

Infrastructure

ML Development

Pre-trained APIs

AI Solutions

Module 1 — Recap ✅

Infrastructure ML Development Pre-trained APIs AI Solutions

📺 Reinforce Tonight on Skills Boost

Module 2 → AI on Google Cloud (~5 min)
Module 2 → AI infrastructure (~7 min)

~12 min · skim if confident

MODULE 02

Pick a Path 🛤️

Four choices. Each fits a different team and problem.

4 Paths to AI 🛤️

📞

Pre-trained API

Call Google's models. Easiest.

📊

BigQuery ML

SQL skills only. Data in BQ.

🤖

AutoML

Bring data, no model code.

🧪

Custom Training

Full control. ML team required. Hardest.

Pick by team skill, not "power"SQL team → BQ ML. Analyst team → AutoML. ML engineers → Vertex AI Custom.

Try It · Module 2

Pick the Right Path 🛤️

Drop each real-world scenario into the cheapest path that fits.

OCR receipts · no ML team

Predict churn · 80M rows in BigQuery · SQL team

Image classifier · labelled photos · no coders

Novel research model · 3 ML PhDs on staff

Sentiment on Bahasa Malay reviews

Time-series forecast · data already in BQ

Tabular fraud detector · analysts have data

📞 Pre-trained API

📊 BigQuery ML

🤖 AutoML

🧪 Custom Training

Trigger words: "no ML team" → API. "in BigQuery + SQL" → BQ ML. "labelled, no coders" → AutoML. "ML PhDs / novel" → Custom.

Knowledge Check · Module 2

5 Quick Wins 🧪

Q1Team has zero ML experience. Needs to OCR receipts. Best path?

Vision API (pre-trained)

Custom TensorFlow model

Train from scratch

Build with Vertex AI custom training

Q250M rows in BigQuery. SQL skills available. Predict churn. Pick:

Export to CSV → Python

Hire ML team first

BigQuery ML

Vision API

Q3Have labelled image data but no ML team. Best path?

Custom training

Pre-trained Vision API

AutoML Vision

BigQuery ML

Q4Unique research problem with a strong ML team. Best path?

Pre-trained API

AutoML

BigQuery ML

Vertex AI Custom Training

Q5Order the 4 paths from EASIEST to HARDEST:

Pre-trained API → BQ ML → AutoML → Custom

Custom → AutoML → BQ ML → Pre-trained API

AutoML → Pre-trained API → Custom → BQ ML

BQ ML → Custom → AutoML → Pre-trained API

Module 2 — Recap ✅

Pre-trained API BigQuery ML AutoML Custom Training

📺 Reinforce Tonight on Skills Boost

Module 4 → AI development options (~5 min)
Module 4 → Vertex AI + AutoML + Pre-trained APIs + Custom training (~19 min)
🎯 Lab in class → Entity & Sentiment Analysis with Natural Language API (~45 min)

~70 min · we do this one together — it's the easiest entry

🎯 Live Lab · ~45 min

Lab — Entity & Sentiment Analysis 💬

Skills Boost lab: "Entity and Sentiment Analysis with the Natural Language API". Open Skills Boost → search for it → click Start Lab.

🔧

What you'll do

Enable the NL API in a temp GCP project
Run gcloud ml language analyze-entities on a Wikipedia article
Run sentiment analysis on a customer review
Read the JSON output — entities, sentiment, salience

✅

What success looks like

JSON with entities[] populated
Sentiment score between −1 and +1
Each entity has a salience score
Lab marks you ✓ green at the end

If you get stuckDon't refresh Cloud Shell — your credentials reset. If the terminal hangs, type gcloud auth list to confirm you're signed in as the temp lab user.

Goal: see real entities & sentiment scores come out of a one-line API call. That's it. Don't overthink the JSON.

MODULE 03

BigQuery ML 🪄

Train ML models with SQL. No Python. No notebooks.

3 SQL Statements. That's It. ✨

CREATE MODEL my.churn_model OPTIONS(model_type='LOGISTIC_REG') AS SELECT * FROM my.training_data;

SELECT * FROM ML.EVALUATE(MODEL my.churn_model);

SELECT * FROM ML.PREDICT(MODEL my.churn_model, ...);

The whole magicTrain → Evaluate → Predict. All in SQL. All in BigQuery. Predictions plug straight into Looker Studio.

Try It · Module 3

Tap Each Highlighted Keyword 🧪

Click any yellow keyword in this real BQ ML statement. The explainer underneath updates. +2⭐ per new keyword discovered.

CREATE OR REPLACE MODEL proj.ds.churn_model OPTIONS(model_type='LOGISTIC_REG', input_label_cols=['churned']) AS SELECT * EXCEPT(customer_id) FROM proj.ds.customer_features WHERE split = 'TRAIN';

Click a yellow keyword above to see what it does.

After tapping every keyword you've seen the anatomy of a working BQ ML statement. The exam tests these exact pieces.

Knowledge Check · Module 3

5 SQL Wins 🧪

Q1Which SQL function gives predictions?

ML.EVALUATE

ML.PREDICT

ML.TRAIN

ML.RUN

Q2BigQuery ML's biggest advantage?

Best accuracy

Cheapest cost

No data movement — model trains where data lives

Easiest GPU access

Q3Which function evaluates model performance on a hold-out set?

ML.EVALUATE

ML.PREDICT

CREATE MODEL

ML.CHECK

Q4To classify "churned: yes/no", which model_type?

LINEAR_REG

LOGISTIC_REG

KMEANS

ARIMA_PLUS

Q5For monthly sales time-series forecasting, which model_type?

LOGISTIC_REG

KMEANS

ARIMA_PLUS

BOOSTED_TREE_CLASSIFIER

Module 3 — Recap ✅

CREATE MODEL ML.EVALUATE ML.PREDICT LOGISTIC_REG

📺 Reinforce Tonight on Skills Boost

Module 2 → BigQuery ML (~6 min)
🏠 Lab at home → Predict Visitor Purchases with BigQuery ML (~60 min)

~66 min · the lab walks you through every SQL statement we covered. Easiest to do at your own pace tonight.

MODULE 04

Vertex AI 🏭

One platform for everything ML. Recently rebranded to Agent Platform.

4 Things to Remember 🧩

🌳

Model Garden

150+ models. One-click try.

🎨

Agent Studio

Gen AI prompt workbench.

🗃️

Model Registry

Version + govern models.

📡

Endpoints

Deploy & serve predictions.

Online vs Batch predictionOnline = low-latency single requests (chatbot). Batch = many rows at once (nightly scoring). Pick by need.

Try It · Module 4

Online or Batch? ⚡📦

Drag each real use case into the right serving mode.

Customer-service chatbot reply

Nightly churn-risk scoring · 50M customers

Real-time fraud detection on each card swipe

Monthly revenue forecast across all branches

In-app product recommendation on page load

Quarterly customer-segment clustering

⚡ Online Prediction

📦 Batch Prediction

Trigger words: "real-time / on click / per call" → Online. "Nightly / monthly / many rows" → Batch.

Knowledge Check · Module 4

5 Quick Wins 🧪

Q1Which component versions and governs models before serving?

Endpoints

Model Registry

Agent Studio

Workbench

Q2Score 50M customers once a month. Best serving mode?

Online prediction

Batch prediction

Real-time stream

Manual prediction

Q3Need low-latency chatbot responses. Best serving mode?

Online prediction via Endpoint

Batch prediction

Overnight scoring job

Manual CSV upload

Q4Where do you browse 150+ models in one click?

Cloud Storage

BigQuery

Endpoints

Model Garden

Q5Vertex AI is rebranding to:

Cloud AI

Gemini Studio

Agent Platform

AI Workshop

Module 4 — Recap ✅

Model Garden Agent Studio Model Registry Endpoints Online vs Batch

📺 Reinforce Tonight on Skills Boost

Module 3 → Generative AI on Google Cloud + Idea to app (~14 min)
Module 3 → Deployment & model tuning (~8 min)
Module 5 → Model serving (~3 min)
🎯 Lab in class → Get Started with Agent Studio (~60 min)

~85 min · we do this together — it's the highlight of the whole course

MODULE 05

MLOps Basics 🔄

How models survive in production. Three ideas — that's it.

3 Ideas You'll Be Asked About 🎯

1. MLOps Maturity LevelsLevel 0 = manual. Level 1 = automated pipelines. Level 2 = CI/CD. Most teams are at 0. Go to 1.

2. DriftLive data slowly drifts away from training data. Model accuracy quietly drops. Monitor for it.

3. Continuous trainingWhen drift hits, auto-retrain on fresh data. Deploy via canary. Loop.

Try It · Module 5

Honest Self-Check — Where Are You? 🪞

Click the level that sounds like your team. The card reveals what you have and what you're missing.

LEVEL

0

Manual everything

LEVEL

1

Pipelines automated

LEVEL

2

CI/CD for pipelines

Most enterprises = Level 0. Moving to Level 1 unlocks 80% of the value. Don't chase Level 2 unless you ship a lot of models.

Knowledge Check · Module 5

5 Final Wins 🧪

Q1Model accuracy drops 87% → 62% over three months. Most likely cause?

The model "broke"

Data drift — live behaviour changed

BigQuery got slower

Endpoint scaling failure

Q2Team at MLOps Level 0. Highest-leverage upgrade?

Buy more GPUs

Automate the training pipeline (move to Level 1)

Hire 10 more data scientists

Switch cloud providers

Q3Level 2 MLOps maturity means:

Manual deployment from a notebook

Automated retraining only

Full CI/CD for ML pipelines themselves

Real-time model inference

Q4Training data differs from live serving data from day 1. That's called:

Training-serving skew

Concept drift

Overfitting

Hallucination

Q5Drift detected → auto-retrain on fresh data → canary deploy. That's:

Fine-tuning

Manual deployment

A/B testing

Continuous training

Module 5 — Recap ✅

Level 0 / 1 / 2 Drift Continuous training

📺 Reinforce Tonight on Skills Boost

Module 5 → ML workflow + Data preparation + Model development (~13 min)
Module 5 → MLOps and workflow automation (~6 min)
Module 5 → How a machine learns (optional but recommended) (~11 min)
🏠 Lab at home → Agent Platform: Predicting Loan Risk with AutoML (~60 min)

~90 min · do this one tonight at your own pace — it earns the official Google Cloud badge

You've got the whole Day 4 framework. 🎯

🎯 Closing Lab · ~60 min

Final Lab — Get Started with Agent Studio 🎨

Last live lab of the day. Skills Boost: "Get Started with Agent Studio". The headline lab of the whole course — leave on a high.

🔧

What you'll do

Open Agent Studio in the GCP Console
Pick a Gemini model (Pro or Flash)
Write a prompt → adjust temperature, top-p, max tokens
Test multimodal input (paste an image or PDF)
Save the prompt as a reusable template

✅

What success looks like

Prompt returns coherent text in the response pane
Lower temperature → more deterministic output
Same prompt with image input returns image-grounded answer
Template appears in your saved prompts list

Pro tip mid-labTry the SAME prompt at temperature 0.0 vs 1.5 to feel the difference. That's the moment "sampling parameters" stops being abstract.

Goal: leave today knowing you could build a Gen AI prototype tomorrow. Then wrap-up + homework, and we're done.

Group Activity · 15 min

Think · Pair · Share 💭

Pair up with someone nearby. Pick one prompt. Discuss 5 min. Share with the room.

📊

Which Path?

For your workplace's biggest data problem — Pre-trained API, BQ ML, AutoML, or Custom? Why?

🔄

One Pipeline

Sketch a continuous training pipeline for one real use case. What triggers retraining?

📡

Online or Batch?

Pick three live use cases at your work. Which need online prediction, which batch? Why?

No wrong answers. Goal: connect today's stack to real problems you actually have.

Screenshot This 📸

Your Exam Cheat Sheet 📋

All 20 concepts on one screen. Take a photo. Review before the exam.

🥞 The Stack

InfrastructureML DevelopmentPre-trained APIsAI Solutions

🛤️ Pick a Path

Pre-trained APIBigQuery MLAutoMLCustom Training

🪄 BigQuery ML

CREATE MODELML.EVALUATEML.PREDICTLOGISTIC_REG

🏭 Vertex AI / Agent Platform

Model GardenAgent StudioModel RegistryEndpointsOnline vs Batch

🔄 MLOps

Level 0 / 1 / 2DriftContinuous training

Match these 20 keywords to their meanings and you're 80% to a pass. Easy.

Final Practice · 10 Mixed Questions

Exam Simulation 🎯

Mixed-module questions. Same shape as the real exam. Answers shuffled.

F1Vertex AI sits in which layer?

ML Development

Infrastructure

Pre-trained APIs

AI Solutions

F250M rows in BigQuery, SQL team, predict churn. Pick:

BigQuery ML

Custom training

Vision API

Export to Python

F3SQL function for predictions:

ML.PREDICT

ML.EVALUATE

ML.TRAIN

CREATE MODEL

F4Score 50M customers monthly. Best mode:

Batch prediction

Online prediction

Stream processing

Manual CSV

F5Versions and governs models before serving:

Model Registry

Endpoints

Agent Studio

Workbench

F6Silent accuracy drop over 3 months. Cause?

Data drift

Model broke

BigQuery slow

Endpoint failure

F7MLOps Level 0 team's biggest upgrade:

Automate the pipeline (Level 1)

Buy more GPUs

Hire more engineers

Switch clouds

F8Vertex AI is rebranding to:

Agent Platform

Cloud AI

Gemini Studio

AI Workshop

F9Training data ≠ live data from day 1 is called:

Training-serving skew

Concept drift

Overfitting

Hallucination

F10For monthly sales time-series forecasting:

ARIMA_PLUS

LOGISTIC_REG

KMEANS

LINEAR_REG (classification)

7+ correct → exam ready. 5-6 → solid, finish Skills Boost tonight. Under 5 → no worries, the labs will lock it in.

Exam Day — Know What to Expect ⏰

📋

Format

50–60 multiple choice questions

⏱️

Time

90 minutes · ~90 sec per question

🎯

Pass Mark

~75% · aim for 80% safe

🧠 Test-taking tips

1. First read — answer everything you know fast. Flag hard ones.
2. Second pass — return to flagged with fresh eyes.
3. Eliminate — wrong answers are usually obvious. Cross them off.
4. Trust patterns — "drift" answers production-drop questions. "Batch" answers high-volume monthly questions. "ML Development" hosts Vertex AI / BigQuery ML.
5. Never leave blank — no negative marking. Guess if you must.

90 minutes for 60 questions = generous. You've prepared. Breathe. You'll do well.

Tonight 📝

Do the Labs — Earn the Badge 🏅

You've got the framework. Tonight you click through the official labs.

🔗 Introduction to AI and Machine Learning on Google Cloud

skills.google/course_templates/593

6 modules · 4 hands-on labs · Google Cloud badge on completion

🧪

Lab 1

Visitor Purchases with BQ ML

🎨

Lab 2

Get Started with Agent Studio

💬

Lab 3

NL API — Sentiment Analysis

Doing the labs is what gets the framework to stick. One hour tonight = real confidence tomorrow.

DAY 4 COMPLETE

You're ready. 🚀

Two days. Two frameworks. 45 exam concepts. Tomorrow you take the exam — and pass. Rest tonight. Hydrate. We've done the hard part.