AI engineer roadmap

The roadmap at a glance

1. 1Python and developer tools
2. 2Building with LLMs
3. 3Math for AI
4. 4Machine learning
5. 5Deep learning
6. 6Transformers
7. 7Modern LLM engineering
8. 8Software engineering
9. 9APIs
10. 10Model Context Protocol (MCP)
11. 11Agent harnesses
12. 12Shipping and getting hired

1. 1

   Stage 1: Python and developer tools

   24 lessons · about 12 hours

   Everything after this is Python calling APIs, so start here. It assumes nothing, not even what a variable is, and you send your first request to a language model before it ends.

   • Python core
   • Working with data
   • Developer tools
   • Talking to APIs

   You'll build: Ad-copy reviewer. Combine files, functions, JSON and an LLM call into one small tool. Free

   All 24 lessons

   1. Your first Python program, step by step Free
   2. Python variables explained for beginners Free
   3. Python numbers: integers, floats and rounding Free
   4. Python strings and f-strings explained Free
   5. Python if, elif and else statements Free
   6. Python lists: indexing, slicing and changing Free
   7. Python for loops and while loops Free
   8. Python dictionaries explained with examples Free
   9. Python functions: arguments and return values Free
   10. Python list comprehensions explained Free
   11. Python errors, tracebacks and try/except Free
   12. Python modules and imports explained Free
   13. Python classes and objects for beginners Free
   14. Working with JSON in Python Free
   15. Read and write files and CSV in Python Free
   16. NumPy arrays for beginners Free
   17. Terminal commands for beginners Free
   18. Git basics for beginners: commits and undo Free
   19. pip and venv: Python packages and environments Free
   20. How APIs work: HTTP requests and responses Free
   21. How to call an API in Python Free
   22. What is an API key? Keeping keys secret Free
   23. Your first LLM API call in Python Free
   24. Python project: build an AI ad-copy reviewer Free

   Related: Free Python for AI course

2. 2

   Stage 2: Building with LLMs

   31 lessons · about 17 hours

   This is most of the day-to-day work: prompts, embeddings and retrieval, tools, agents, evals, cost and security. None of it needs math.

   • Prompting
   • Context & retrieval
   • Tools & agents
   • Quality & production

   You'll build: Research agent. A marketing research agent with retrieval, tools and evals.

   All 31 lessons

   1. What is an LLM? How language models work Free
   2. LLM tokens and context windows explained
   3. Prompt engineering basics, with examples
   4. What is a system prompt? Roles explained
   5. Few-shot prompting explained with examples
   6. LLM structured output: getting reliable JSON
   7. Prompt chaining explained with examples
   8. How to choose an LLM: capability, speed, cost
   9. Reasoning models: how thinking effort works
   10. How to stream LLM responses in Python
   11. How LLM memory works: chat history explained
   12. Sending images and PDFs to an LLM
   13. What are embeddings? Explained with examples
   14. Similarity search with embeddings explained
   15. Chunking for RAG: how to split documents
   16. What is a vector database? How it works
   17. What is RAG? Retrieval-augmented generation
   18. Improve RAG: hybrid search and reranking
   19. Why LLMs hallucinate, and how to reduce it
   20. LLM function calling and tool use explained
   21. How AI agents work: the agent loop explained
   22. What is MCP? Model Context Protocol explained
   23. AI agent design patterns explained
   24. Using LLMs in automations and workflows
   25. What is prompt injection? How to limit it
   26. LLM evals: how to measure output quality
   27. LLM as a judge: grading output with rubrics
   28. Reduce LLM cost: prompt caching and batching
   29. LLM observability: logging and monitoring
   30. How to deploy an LLM app to production
   31. Build an AI research agent: a full project

   Related: Free prompt engineering course · Free RAG course · Free AI agents course · Prompt engineering interview questions and answers · RAG interview questions and answers · AI agent interview questions and answers

3. 3

   Stage 3: Math for AI

   29 lessons · about 15 hours

   Only what models actually run on: vectors and matrices, derivatives and gradients, probability and cross-entropy. You can put it off while you're building apps, but the next four stages build on it.

   • Linear algebra
   • Calculus
   • Probability & statistics
   • Information

   You'll build: Cross-entropy and KL divergence. The loss function language models are trained on.

   All 29 lessons

   1. What is a vector? Linear algebra for ML Free
   2. Vector addition, scaling and length explained
   3. Dot product and cosine similarity explained
   4. What is a matrix? Matrices in machine learning
   5. Matrix multiplication explained step by step
   6. Linear transformations: what a matrix does
   7. Matrix transpose, identity and inverse
   8. What is a tensor? Explained for deep learning
   9. Eigenvectors explained with intuition
   10. PCA and SVD explained with intuition
   11. Functions and graphs for machine learning
   12. Exponentials and logarithms explained
   13. What is a derivative? Explained with slopes
   14. Derivative rules: power, sum and product
   15. Chain rule explained for backpropagation
   16. Partial derivatives explained simply
   17. What is a gradient in machine learning?
   18. Gradient descent explained: the math
   19. Probability basics for machine learning
   20. Conditional probability and Bayes' rule
   21. Probability distributions explained
   22. Mean, variance and standard deviation
   23. The normal distribution explained
   24. Standard error and confidence intervals
   25. A/B testing and statistical significance
   26. Maximum likelihood estimation explained
   27. Softmax function explained with examples
   28. Entropy in machine learning explained
   29. Cross-entropy loss and KL divergence explained

   Related: Gradient descent, visualized

4. 4

   Stage 4: Machine learning

   24 lessons · about 14 hours

   How models learn from data, and how to tell whether one really works. Habits from here, like holding back test data and measuring precision and recall, carry over to everything you build later.

   • Foundations
   • Models
   • Evaluation
   • Unsupervised

   You'll build: Churn prediction. Predict which customers will leave, and measure it properly.

   All 24 lessons

   1. What is machine learning? A clear introduction Free
   2. Supervised learning: features and labels
   3. Data preparation for ML: clean, encode, scale
   4. Feature engineering explained with examples
   5. Train, validation and test sets explained
   6. Linear regression explained with Python
   7. Loss functions in machine learning explained
   8. Gradient descent in Python, step by step
   9. What is the learning rate in machine learning?
   10. Logistic regression explained with Python
   11. Multiclass classification with softmax
   12. Decision trees explained with examples
   13. Random forests and gradient boosting explained
   14. K-nearest neighbors (KNN) explained
   15. Overfitting and underfitting explained
   16. Regularization in machine learning: L1 and L2
   17. The bias-variance tradeoff explained
   18. Precision, recall and F1 score explained
   19. ROC curves and AUC explained
   20. K-fold cross-validation explained
   21. K-means clustering explained
   22. Dimensionality reduction: PCA and t-SNE
   23. Scikit-learn tutorial: fit, predict, evaluate
   24. Customer churn prediction in Python: a project

   Related: ROC curve, precision and recall, explained interactively · Overfitting, visualized · Machine learning interview questions and answers

5. 5

   Stage 5: Deep learning

   23 lessons · about 15 hours

   Neurons, backpropagation and the craft of getting a network to train, including an autograd engine you write yourself, then PyTorch.

   • Neural networks
   • Training
   • Architectures
   • Tooling

   You'll build: Train a classifier. Build, train and debug a network end to end.

   All 23 lessons

   1. What is an artificial neuron? Explained simply Free
   2. Activation functions: ReLU, sigmoid and GELU
   3. Multi-layer perceptrons (MLP) explained
   4. The forward pass in neural networks explained
   5. Computation graphs in deep learning
   6. Backpropagation explained step by step
   7. Build an autograd engine in Python (micrograd)
   8. Weight initialization in neural networks
   9. The neural network training loop explained
   10. Mini-batch gradient descent and SGD explained
   11. Optimizers explained: SGD, momentum and Adam
   12. Batch norm and layer norm explained
   13. Dropout in neural networks explained
   14. Vanishing and exploding gradients explained
   15. How to debug neural network training
   16. Residual and skip connections explained
   17. Convolutional neural networks (CNN) explained
   18. Recurrent neural networks (RNN) explained
   19. Embedding layers in neural networks explained
   20. PyTorch tensors tutorial for beginners
   21. PyTorch autograd and nn.Module tutorial
   22. Why deep learning needs GPUs, explained
   23. Build and train a neural network classifier

   Related: Neural network playground: train one in your browser · Backpropagation, step by step · Deep learning interview questions and answers

6. 6

   Stage 6: Transformers

   21 lessons · about 14 hours

   Attention from the intuition up, then a small GPT you write and train yourself. After this, a language model is no longer a black box.

   • Language modeling
   • Attention
   • The transformer
   • Scale & inference

   You'll build: Build a tiny GPT. Write and train a small GPT from scratch.

   All 21 lessons

   1. Language models: predicting the next token Free
   2. BPE tokenization explained step by step
   3. Build a bigram language model in Python
   4. Temperature, top-k and top-p explained
   5. Neural language models: beyond counting
   6. Attention in transformers: the intuition
   7. Query, key and value in attention explained
   8. Scaled dot-product attention, line by line
   9. Causal attention masks in transformers
   10. Multi-head attention explained
   11. Positional encoding in transformers explained
   12. The transformer block explained, part by part
   13. GPT architecture explained
   14. Build a GPT from scratch in Python
   15. BERT and embedding models explained
   16. How LLMs are trained: pretraining at scale
   17. LLM scaling laws explained
   18. KV cache in LLM inference explained
   19. Long context in LLMs: efficient attention
   20. Mixture of experts (MoE) explained
   21. Multimodal transformers: images and audio

   Related: Build an LLM from scratch: a free course · How a transformer works, with a real GPT you can run · Attention in transformers, explained interactively · Temperature, top-k and top-p, explained visually · LLM interview questions and answers

7. 7

   Stage 7: Modern LLM engineering

   22 lessons · about 16 hours

   How a raw model becomes an assistant, and how to choose between a better prompt, RAG and fine-tuning. It also covers running models yourself, benchmarks and safety.

   • Post-training
   • Adapting models
   • Inference
   • Systems & safety

   You'll build: A tuned, evaluated assistant. Fine-tune, evaluate and deploy a specialist assistant.

   All 22 lessons

   1. Base model vs chat model: what's different Free
   2. Supervised fine-tuning (SFT) explained
   3. Instruction and preference data for LLMs
   4. RLHF explained: learning from human feedback
   5. DPO explained: direct preference optimization
   6. What is constitutional AI? Explained
   7. How reasoning models are trained
   8. Prompting vs RAG vs fine-tuning: how to choose
   9. LoRA fine-tuning for LLMs explained
   10. How to fine-tune an open LLM: a full project
   11. Knowledge distillation for LLMs explained
   12. LLM quantization explained: GGUF, AWQ and more
   13. LLM serving: batching, throughput and latency
   14. Speculative decoding explained
   15. How to run open LLMs locally
   16. LLM benchmarks explained, and their limits
   17. LLM interpretability explained
   18. AI alignment and safety explained
   19. Multi-agent systems explained
   20. How to evaluate AI agents
   21. How to read AI research papers
   22. Fine-tune and deploy an LLM assistant
8. 8

   Stage 8: Software engineering

   16 lessons · about 12 hours

   What separates a demo from software people can rely on: complexity, concurrency, testing, debugging, design, databases and CI.

   • How computers run code
   • Data structures & algorithms
   • Writing good code
   • Building software

   You'll build: A command-line app. Design, build and test a complete program with a database.

   All 16 lessons

   1. How computers run code, explained for Python Free
   2. Big O notation explained with Python
   3. Python concurrency and asyncio explained
   4. How to profile and speed up Python code
   5. Data structures explained with Python
   6. Recursion, trees and graphs in Python
   7. Sorting, searching and dynamic programming
   8. Python testing for beginners with pytest
   9. How to debug code systematically
   10. Clean code and refactoring in Python
   11. Python type hints and dataclasses explained
   12. Software design basics: modules and interfaces
   13. SQL basics: queries and joins for beginners
   14. Secure coding basics: injection and secrets
   15. Python project structure, packaging and CI
   16. Build a command-line app in Python
9. 9

   Stage 9: APIs

   13 lessons · about 10 hours

   Most AI products reach people through an API. You build one from the wire up, then add validation, a database, authentication, rate limits and webhooks.

   • HTTP from the wire up
   • Building APIs
   • Security & reliability
   • Design at scale

   You'll build: A complete API. An authenticated, validated, paginated API with tests.

   All 13 lessons

   1. How HTTP works: requests and headers explained Free
   2. REST vs RPC vs GraphQL vs gRPC explained
   3. Build an API server in Python from scratch
   4. How Flask and FastAPI work, from the inside
   5. REST API design best practices
   6. Validate API requests with JSON Schema
   7. Build a CRUD API in Python with SQLite
   8. API authentication: keys, tokens and OAuth
   9. API rate limiting, retries and idempotency
   10. How webhooks work: signing and retries
   11. OpenAPI explained: documenting your API
   12. Running APIs in production: a practical guide
   13. Build a complete REST API in Python
10. 10

    Stage 10: Model Context Protocol (MCP)

    14 lessons · about 11 hours

    The standard way to connect models to tools and data. You go from the wire format to a production server with OAuth, and write the client side too.

    • Protocol foundations
    • Building servers
    • Clients and hosts
    • Production MCP

    You'll build: A production MCP server. A complete, tested server for the practice shop.

    All 14 lessons

    1. JSON-RPC 2.0 explained with Python Free
    2. MCP protocol versions and stateless requests
    3. MCP discovery and capabilities explained
    4. MCP stdio transport explained
    5. MCP Streamable HTTP transport explained
    6. Build an MCP server: tools in Python
    7. MCP resources and prompts explained
    8. MCP progress notifications and subscriptions
    9. Build an MCP client in Python
    10. MCP hosts: connecting servers to a model
    11. MCP OAuth: securing remote servers
    12. MCP security: tool poisoning and rug pulls
    13. How to test and deploy an MCP server
    14. Build a production MCP server: a full project
11. 11

    Stage 11: Agent harnesses

    14 lessons · about 10 hours

    The software around a model that turns it into a capable, safe agent: tools, permissions, context management, subagents, tracing and guardrails.

    • Harness foundations
    • Tools & environment
    • Context & control
    • Scaling & shipping

    You'll build: Build a coding agent. A harness that edits a project, runs its tests and stops when they pass.

    All 14 lessons

    1. What is an agent harness? Explained Free
    2. Build the core loop of an AI agent
    3. Build a tool registry for AI agents
    4. Build file and search tools for AI agents
    5. Run agent code safely: sandboxes and limits
    6. AI agent permissions and approval prompts
    7. Agent hooks and skills explained
    8. Agent context management and compaction
    9. AI agent memory and planning for long tasks
    10. What are subagents? How to use them
    11. Agent streaming, interrupts and sessions
    12. Build an evaluation harness for AI agents
    13. AI agent tracing, budgets and guardrails
    14. Build a coding agent in Python
12. 12

    Stage 12: Shipping and getting hired

    13 lessons · about 10 hours

    System design, deployment, CI/CD, incidents, working in a team and preparing for interviews. This is where you put it all into production.

    • System design
    • Shipping
    • Working in a team
    • Career

    You'll build: Ship a production AI service. An AI-backed service with validation, caching, rate limits, fallbacks, logging and evals.

    All 13 lessons

    1. System design fundamentals for beginners Free
    2. Caching, queues and background jobs explained
    3. AI system design: latency, cost and fallbacks
    4. Docker containers and deployment explained
    5. CI/CD pipelines explained for beginners
    6. SLOs, error budgets and incident response
    7. Git for teams: pull requests and rebasing
    8. Code review best practices
    9. How to read an unfamiliar codebase
    10. How to use AI coding tools well
    11. Technical writing for engineers: a guide
    12. AI engineer interviews and portfolio guide
    13. Ship a production AI service: a full project

    Related: AI engineer interview questions and answers