# What You Can Build with Machine Learning

Machine learning is a computer working out a rule from past examples, the way a spam filter, a delivery estimate or a fraud check does, and this lesson shows where it's used, what it isn't (no magic, no promises, and in this course no chatbots), and the projects you'll build.

Novice · AI Basics · Python · Updated November 19, 2026

By Gregor Dzierzon

## After this lesson you can

- Name everyday guesses a model learns from past examples, such as a spam filter, a delivery estimate and a fraud check
- Explain what machine learning isn't, no magic and no promises, and why chatbots and large language models are outside this course

## Key takeaways

- Machine learning is a computer working out a rule from past examples instead of a person writing it. The rule it learns is a model.
- Spam filters, delivery-time estimates, fraud checks and "customers like you" lists are all guesses a model learned from examples.
- A model only knows the examples it learned from. It guesses, it can be wrong, and it never knows why the world works the way it does.
- Chatbots are built with large language models, a kind of machine learning this course doesn't cover. It teaches guesses from tables of data.
- You'll build five projects, from judging when machine learning fits to explaining a model's groups, then a model of your own as the final capstone.

Your email app moves a message to spam before you see it. A delivery app says your order will arrive in 25 minutes. Your bank texts you about a card payment that looks odd. Nobody sat down and wrote a rule for that one email, that one trip or that one payment. Each is a guess a program learned from past examples, and that's what this course teaches you to build, test honestly and explain.

## A rule that keeps breaking

Start with the delivery estimate, the way a small shop would. Pine Ridge Delivery writes a rule: 10 minutes to get going, plus 3 minutes for every mile. Then the real trips come in:

| The trip | The rule's guess | How long it took |
|---|---|---|
| 2 miles, a dry evening | 16 minutes | 15 minutes |
| 2 miles, in the rain | 16 minutes | 21 minutes |
| 2 miles, two stops before yours | 16 minutes | 24 minutes |

So the shop adds to the rule: more time in the rain, more for every stop. How much more? Five minutes for rain, or eight? Someone has to guess every number, and every new problem, such as rush hour or a new neighborhood, needs another rule and another guess.

## Learning the rule from examples

The shop already has the answer to those questions in its records: 500 past deliveries, each with its distance, its stops, the weather and how many minutes it really took. Instead of a person guessing the numbers, a program can work them out from those examples. That is **machine learning**: a computer finding a rule from examples instead of being told it.

The rule it learns is called a **model**. Learning it is called **training**, and the examples it learns from are its **training data**. A model's guess for a new delivery is called a *prediction*, the word your textbook and scikit-learn's documentation use.

You meet models every day:

| What it guesses | The examples it learned from |
|---|---|
| Is this email spam? | Thousands of emails people marked as spam or not spam |
| How long will this delivery take? | Past deliveries with their distance, stops, weather and real time |
| Which books might this reader like? | What readers with similar borrowing borrowed next: a "customers like you" list |
| Is this card payment fraud? | Past payments, with the few that turned out to be fraud |

The delivery estimate guesses a number (minutes), while the spam filter and the fraud check guess a category (spam or not spam, fraud or not). *Kinds of Machine Learning* names these kinds properly.

## What machine learning isn't

It isn't magic. A model can only find a pattern that's in its examples. If the shop's records never noted the rain, no model will learn what rain does, and *When Machine Learning Is the Wrong Tool* shows the problems where a plain rule or a database query does the job better.

Its guesses aren't promises. A model is wrong some of the time, and how often is something you measure on examples it never saw, which is much of what this course teaches. It also doesn't know *why*: it can learn that rainy trips take longer, but not what rain does to traffic.

And it isn't a chatbot. Chatbots that write text are built with machine learning too: **large language models**, very large models trained on huge amounts of text. This course doesn't cover chatbots or large language models. It teaches the machine learning behind everyday guesses from tables of data: a number, a category, or groups of things that are alike, with scikit-learn, a Python package for it, and with straight lines, decision trees and other methods you can understand and explain, not neural networks.

## What you'll build in this course

Each module ends with a project, a **mini-capstone**, with its own starter files and our own made-up data.

- **Is Machine Learning the Right Tool?** (Module 1). A one-page judgment of six problems: machine learning or not, which kind, what it would learn from, the simplest guess to beat, and one way it could go wrong.

  ![A one-page worksheet of six problems, from adding sales tax to a pizza order to grouping a store's shoppers, each with a space for the verdict, machine learning or not, and for the kind of guess, the simplest guess to beat and one risk](https://fsdtrainingmedia.blob.core.windows.net/public/images/2026/11/b3f424c14f8f4d5caf6fc555c035cb54.webp)

- **Bike Share Forecast** (Module 2). A model that predicts a bike share's daily rentals from the weather, with its error in rides and a chart of where it misses.

  ![A year of a bike share's daily rentals as dots, the model's guesses as a line through them, and the days it missed by the most circled, with its typical error written in rides](https://fsdtrainingmedia.blob.core.windows.net/public/images/2026/11/f599bb3974ce48c2a4ea493b5eeb0c14.webp)

- **Library Renewals** (Module 3). A model that flags library members who may not renew, with its mistakes counted and a cutoff chosen for a reminder email.

  ![Library members sorted by the model's chance that they won't renew, a cutoff line choosing who gets a reminder email, and a small table counting the model's two kinds of mistakes](https://fsdtrainingmedia.blob.core.windows.net/public/images/2026/11/7850188e549f480b9350e4aa7c0469d4.webp)

- **The Model That Only Seemed to Work** (Module 4). A used-car price model that scored too well, with every leak found, fixed and honestly rescored.

  ![A used-car price model's score dropping from nearly perfect to a believable one as three leaks are found and fixed: a column that held the answer, cars listed twice, and data prepared before the split](https://fsdtrainingmedia.blob.core.windows.net/public/images/2026/11/08808910f382485d8bb356678b60dd8d.webp)

- **Shopper Groups** (Module 5). A grocery's shoppers sorted into groups, each described in plain words, with what the grouping can't tell you.

  ![A grocery's shoppers sorted into five groups of dots, each group with a short plain-words description card, and a note on what the grouping can't tell you](https://fsdtrainingmedia.blob.core.windows.net/public/images/2026/11/97018388f628400098f782d5dd00d696.webp)

- **Your own model** (the final capstone). A question you choose, a model compared, measured and explained, with a short card that says where it shouldn't be trusted.

  ![The final capstone's parts in order: a question you choose, the data prepared, a simple guess and two models compared, one honest test score, and a model card that says where the model shouldn't be trusted](https://fsdtrainingmedia.blob.core.windows.net/public/images/2026/11/16e5e439ba8e4f459c472d3d9bdbb823.webp)

## Check yourself

2 questions to try on the lesson page.

## So what

A model is a rule learned from examples, honest only about the examples it saw, and this course teaches you to build one, measure it and say plainly what it can and can't do.

## Sources

- scikit-learn 1.9.1, [Getting Started](https://scikit-learn.org/stable/getting_started.html): scikit-learn is an open source machine learning library that supports supervised and unsupervised learning; an estimator is fitted to data, then predicts target values for new data.
- scikit-learn, [Glossary of Common Terms and API Elements](https://scikit-learn.org/stable/glossary.html): *sample* (a single example, also called an instance or observation), *target* (the value passed as `y` to `fit`, also known as the label), and *predict* (makes a prediction for each sample).
- scikit-learn 1.9.1, [Supervised learning](https://scikit-learn.org/stable/supervised_learning.html) and [Unsupervised learning](https://scikit-learn.org/stable/unsupervised_learning.html), User Guide: the library's methods, from linear models and decision trees to clustering, and its *Neural network models (supervised)*, which this course leaves out.

## Key terms

- [Machine Learning](https://www.ask-codey.com/glossary/machine-learning)
- [Model](https://www.ask-codey.com/glossary/model-ai)
- [Training Data](https://www.ask-codey.com/glossary/training-data)
- [Large Language Model](https://www.ask-codey.com/glossary/large-language-model)

## Related

- [How This Course Works](https://www.ask-codey.com/lessons/ml-how-this-course-works)
- [Guessing from Examples](https://www.ask-codey.com/lessons/ml-guessing-from-examples)
- [When Machine Learning Is the Wrong Tool](https://www.ask-codey.com/lessons/ml-when-ml-is-the-wrong-tool)

Source: https://www.ask-codey.com/lessons/ml-what-you-can-build
