Learn AI · Level 4 · Frontier view · step 16 of 16

From AI literacy to building AI: a roadmap

Updated 29 Oct 2026 · Reviewed every quarter

The Learn AI path teaches you to understand AI. This roadmap is for the next step: building it. Six stages take you from your first line of Python to an AI app running in production, each with what you'll be able to do, the best courses (most of them free) and a small project to prove it. It's a real commitment, so the times are honest.

Stage 0: you've done the Learn AI path. You know what AI is, how models learn and the words people use. The path is for understanding AI; this roadmap is for building it.

1Python6–10 weeks
2Maths6–8 weeks
3Classical ML2–3 months
4Deep learning2–3 months
5LLM apps1–2 months
6Production1–2 months

Be honest about the time. At 8–10 hours a week, the six stages take roughly 8 to 13 months, longer if you're new to maths. You can build a first small project within 2 to 3 months. Being ready for a job usually takes a year or more, and depends on your projects more than on certificates. (Our estimates, not a promise.)

1
Learn Python
6–10 weeks
Almost all AI work is done in Python. You don't need to become a software engineer, but you need to read and write it comfortably.
You'll be able to
- Write scripts with functions, loops and conditions
- Load, clean and save data files
- Find and fix your own bugs, and write simple tests
- Save and share your code with Git and GitHub
Our picks
Build this: Take a messy export, such as a bank statement or a sports table, and write a script that cleans it and prints a summary.
Already code in another language? Do the Kaggle course in an afternoon and move on.
2
The maths you actually need
6–8 weeks, alongside Python
You don't need a maths degree. Three areas matter: linear algebra (how data is stored and transformed), a little calculus (how models learn) and statistics (how you tell if they work).
You'll be able to
- Picture vectors and matrices, and what multiplying them does
- Explain how gradient descent nudges a model towards fewer mistakes
- Read averages, spread, probability and confidence intervals without guessing
Our picks
Build this: Work out two steps of gradient descent by hand for a line through five points, then check your answer in Python.
Building apps on existing models? A light pass is enough. Come back when something stops making sense.
Gradient descent Loss function
3
Classical machine learning
2–3 months
Most AI in business still predicts numbers and categories from tables: fraud, prices, customers about to leave. This is where you learn the whole workflow, and how to test a model honestly.
You'll be able to
- Train a model with scikit-learn and test it on data it has never seen
- Choose between linear models, decision trees and gradient boosting
- Judge a model with precision and recall, and spot overfitting
Our picks
Build this: Predict house prices or which customers will leave, using a public dataset. Write one page on how good the model really is and where it fails.Datasets to try: House prices, Customers who leave (telecom), Card fraud, Spam text messages
Our guide: How a machine learning model is built →Supervised learning Classification vs regression Decision tree Gradient boosting Train-test split Overfitting Precision and recall
4
Deep learning
2–3 months
Neural networks power image recognition, speech and language models. You will train them yourself, and see why GPUs matter.
You'll be able to
- Build and train neural networks in PyTorch
- Fine-tune a pretrained image or text model on your own data
- Explain how a transformer processes text
Our picks
Build this: Fine-tune an image model to tell apart something you care about, such as your plants or local birds, and put it online as a small demo. A free Hugging Face account can host up to two Gradio demos.
No powerful computer needed: Google Colab and Kaggle notebooks give you free GPU time, within usage limits.
Neural network Convolutional neural network (CNN)Recurrent neural network (RNN)Transformer Transfer learning Fine-tuning GPU
5
Building with large language models
1–2 months
This is where most new AI work is: building products on top of existing models, not training them from scratch.
You'll be able to
- Call a model API and get structured, reliable output
- Add your own documents with RAG, and actions with tool use
- Measure quality with a set of test questions before you ship
- Fine-tune a small open model with LoRA when prompting is not enough
Our picks
Build this: A chatbot that answers questions from your own documents and shows its sources, plus 20 test questions you rerun after every change.
Our guide: RAG explained →RAG (retrieval-augmented generation)Vector database Function calling MCP (Model Context Protocol)AI agent LoRA (low-rank adaptation)Evals (AI evaluations)
6
Shipping to production
1–2 months
A model on your laptop helps nobody. Production means it runs reliably, costs what you expect, and you notice when it gets worse.
You'll be able to
- Package and deploy a model or AI app behind an API
- Test code, data and model, and automate releases
- Watch quality, cost and drift after launch
Our picks
Build this: Put your chatbot from stage 5 online, with logging, a spending limit and tests that run automatically on every change.
MLOps Model drift

Which route fits you?

Three common AI jobs, and the stages each one leans on. Everyone needs Python.

Role	1 Python	2 Maths	3 Classical ML	4 Deep learning	5 LLM apps	6 Production
AI engineerBuilds products on existing models: chatbots, agents, AI search.	Core	Light	Light	Optional	Core	Core
Machine learning engineerTrains, deploys and maintains models.	Core	Core	Core	Core	Light	Core
Data scientistFinds answers in data and builds models that predict.	Core	Core	Core	Light	Optional	Optional

Build a portfolio

Finish two or three projects properly, rather than starting ten.
Use real, messy data, not only tutorial datasets.
Write up each project: the problem, what you tried, how you measured it and what failed.
Put the code on GitHub and, where you can, a live demo anyone can try, for example on Hugging Face Spaces.

Common traps

Endless tutorials. Watching course after course without building. Do the stage project before you start the next course.
Skipping evaluation. A demo that works once is not a product. Always test on data the model has not seen.
Chasing every new tool. Frameworks change every few months. Python, statistics and evaluation don't.
Waiting until you feel ready. Start a project when you know about half of what it needs. The rest comes from getting stuck.

How we put this roadmap together

The six stages follow the order most AI engineering and machine learning curricula use: programming and maths first, then classical machine learning, deep learning, building on large language models and running models in production. Most of our picks are free or have free versions; where one is paid, its card says so. We checked each one on its official page. The times are our own estimates for someone studying 8–10 hours a week; people with a coding or maths background will move faster. We review this page every quarter. Spotted something to fix? Tell us.

Learn AI path · Level 4 · Frontier view · step 16 of 16

You've reached the end of the path. Well done!

See all of level 4

← All guides

From AI literacy to building AI: a roadmap

Learn Python

The maths you actually need

Classical machine learning

Deep learning

Building with large language models

Shipping to production

Which route fits you?

Build a portfolio

Common traps

How we put this roadmap together