50 lessons · dependency-ordered
From your first line of Python
to shipping AI.
A free, open-source roadmap through the whole field. Read it here or straight from the Markdown on GitHub — and your progress saves as you go.
- 0
- completed
- 50
- lessons
- 7
- stages
- ~42h
- of reading
Stage 00 · 2 lessons
Orientation
Get the lay of the land before you write a line of code.
Stage 01 · 5 lessons
Python & Data
The tools you'll touch every day — Python, NumPy, pandas, plots.
- Python BasicsVariables, lists, dicts, loops, functions, imports, and how to read the error messages you'll see constantly.60 minbeginner
- NumPyWhy arrays beat lists for numeric work, plus shape, dtype, indexing, broadcasting, and vectorised math.45 minbeginner
- PandasSeries and DataFrames, loading CSVs, selecting and filtering rows, grouping, and handling missing values.50 minbeginner
- Data VisualizationChoosing the right chart type, matplotlib basics, plotting straight from pandas, and reading what a histogram tells you.40 minbeginner
- Real DatasetsWhere to find datasets, loading messy CSVs, fixing types and dates, removing duplicates, spotting outliers, and a reusable cleaning checklist.45 minbeginner
Stage 02 · 3 lessons
Math You Actually Need
Just enough linear algebra, calculus, and statistics — taught through the problems they solve.
- Linear AlgebraVectors and matrices as the data structures behind every model, dot products as weighted sums, and matrix multiplication as batch prediction.50 minbeginner
- CalculusDerivatives as slope, the chain rule, and gradient descent implemented by hand in NumPy — how a model actually learns.50 minintermediate
- Probability & StatisticsDistributions, mean and variance, conditional probability, Bayes' theorem, sampling, and what a p-value actually means.50 minbeginner
Stage 03 · 11 lessons
Classical ML
Regression to random forests: how machines learn from tables of data.
- ML FundamentalsSupervised vs unsupervised learning, the train/validation/test split, and overfitting vs underfitting through the bias-variance tradeoff.45 minbeginner
- RegressionLinear regression from the normal equation to scikit-learn, the MSE/MAE/R² metrics, and Ridge/Lasso regularisation to fight overfitting.50 minbeginner
- ClassificationLogistic regression and the sigmoid, decision boundaries, k-nearest neighbours, and strategies for more than two classes.50 minbeginner
- Model EvaluationThe confusion matrix, precision, recall, F1, ROC-AUC, cross-validation, and why accuracy alone can make a worthless model look great.45 minintermediate
- Feature EngineeringScaling, encoding categoricals, dates, binning, interaction terms, and data leakage — raising a model's score with features, not a new algorithm.45 minintermediate
- Clustering & PCAk-means clustering, choosing k with the elbow method, hierarchical clustering, and PCA as compression with explained variance.45 minintermediate
- Trees & EnsemblesDecision trees and how splits are chosen, random forests as bagging, gradient boosting, and feature importance.50 minintermediate
- First ML ProjectA full guided ML pipeline end to end: problem framing, EDA, cleaning, features, baseline, iteration, evaluation, and writing up results.90 minintermediate
- Explainable AI — Opening the Black BoxWhy a model made a prediction — feature importance, permutation importance, and reading a model you can't see inside.45 minintermediate
- Time-Series ForecastingPredicting what happens next — why time data breaks normal ML, lag features, a proper time-aware split, and honest baselines.50 minintermediate
- Recommender SystemsThe engine behind every 'you might also like' — content-based vs collaborative filtering, similarity, and the cold-start problem.45 minintermediate
Stage 04 · 9 lessons
Deep Learning
Neural networks, from a single neuron up to the transformer.
- Neural NetworksWhat one neuron computes, why nonlinear activations are non-negotiable, how depth builds representations, and the universal approximation intuition.55 minintermediate
- Backprop & TrainingHow the chain rule turns one output error into a gradient for every weight in a network, and how learning rate, epochs and batches shape training.60 minintermediate
- PyTorchTensors, autograd, nn.Module, optimisers, and the canonical training loop — the framework that automates the backprop you just wrote by hand.60 minintermediate
- CNNs & VisionConvolution as a learned filter, stride, padding and pooling, how a CNN's shapes flow layer to layer, and transfer learning with a pretrained backbone.55 minintermediate
- Sequence ModelsWhy order matters, how RNNs process sequences step by step, the vanishing gradient problem, LSTM/GRU, and why attention replaced them.50 minadvanced
- TransformersAttention as a learned lookup over query, key and value, self-attention and multi-head attention, positional encoding, and the encoder/decoder split.70 minadvanced
- Reinforcement LearningLearning from reward instead of labels — agents, states, actions, rewards, and Q-learning taught by making an agent solve a tiny grid world.55 minadvanced
- Object Detection & SegmentationBeyond 'what's in this image?' to 'what's where?' — bounding boxes, IoU, non-max suppression, and how YOLO detects in real time.50 minadvanced
- Generative Adversarial Networks (GANs)Two networks in a duel — a generator faking data and a discriminator catching fakes — and how that adversarial game learns to create.50 minadvanced
Stage 05 · 17 lessons
Generative AI & LLMs
How modern language models actually work, and how to build with them.
- How LLMs WorkTokenization, next-token prediction, pretraining vs post-training, context windows, temperature and sampling, and why models hallucinate.55 minintermediate
- Prompt EngineeringClear instructions, few-shot examples, chain-of-thought, structured output, system prompts, and fixing a failing prompt in documented iterations.45 minbeginner
- EmbeddingsText as vectors, cosine similarity, embedding models, vector databases, and chunking strategy, with a real semantic search built in NumPy.45 minintermediate
- Retrieval-Augmented Generation (RAG)Why retrieval beats stuffing the context window, the ingest-chunk-embed-retrieve-generate pipeline, chunk sizing, and common RAG failure modes.60 minintermediate
- Fine-tuningWhen fine-tuning beats RAG or prompting, full fine-tuning vs LoRA/PEFT, dataset preparation, and evaluating the result, with a real LoRA parameter-count demo.60 minadvanced
- AI AgentsTool use, the reason-act loop, planning, memory, multi-step failure modes, and cost control, with a real non-LLM demo of the loop mechanics.60 minadvanced
- Evals & GuardrailsWhy manual spot-checking doesn't scale, building an eval set, LLM-as-judge and its biases, regression testing, and input/output guardrails.50 minadvanced
- Run Open LLMs Locally with OllamaRun real language models on your own machine — private, free, offline — and call them from Python like an API.45 minintermediate
- Build an Agent HarnessThe scaffolding that turns a raw language model into an agent: the tool loop, parsing, dispatch, and history — built from scratch.60 minadvanced
- Loop EngineeringThe naive agent loop breaks in a dozen ways. The engineering that makes it reliable: budgets, loop detection, context control, and error recovery.55 minadvanced
- Diffusion Models & Image GenerationHow Stable-Diffusion-style models turn noise into images — the forward noising process, learning to denoise, and why it works.55 minadvanced
- Vector DatabasesWhere embeddings live at scale — similarity search, why brute force stops scaling, and how approximate nearest neighbors makes it fast.45 minintermediate
- Structured Outputs & Function CallingGetting reliable JSON out of an LLM — schemas, validation, function/tool calling, and what to do when the model returns something wrong.45 minintermediate
- Model Context Protocol (MCP)The USB-C of AI tools — a standard protocol that lets any agent connect to any tool or data source without custom glue for each one.45 minadvanced
- AI Ethics & Responsible AIBuilding AI that doesn't cause harm — bias and fairness, privacy, transparency, misuse, and the responsibility that comes with shipping models.45 minintermediate
- Speech & Audio AIHow machines hear — turning sound into spectrograms, speech-to-text with Whisper, and text-to-speech, from waveform to model input.45 minintermediate
- Multimodal AI (CLIP & Vision-Language Models)One model, many senses — how CLIP puts images and text in the same space, enabling zero-shot classification and image search.50 minadvanced
Stage 06 · 3 lessons
Ship It
Take a model off your laptop and put it in front of real users.
- MLOps BasicsExperiment tracking, model and data versioning, reproducibility, and drift monitoring — what keeps a shipped model trustworthy after it leaves your notebook.45 minintermediate
- Deploying a ModelWrapping a trained model behind a validated FastAPI endpoint, measuring latency and batching, Dockerizing it, and free hosting options.55 minintermediate
- Portfolio and CareerWhat makes a project worth showing, writing a README that gets read, the real AI job families, and how to keep learning after this roadmap.40 minbeginner
Frequently asked questions
Is this AI roadmap free?
Yes — completely free and open source. All 50 lessons are free to read on the site or straight from the Markdown on GitHub, with no sign-up, paywall, or ads.
How long does it take to learn AI with this roadmap?
The 50 lessons are about 42 hours of reading, plus time to run the code and do the exercises. Most people spread it over a few months of steady study — you can go faster or slower, and your progress saves as you go.
Do I need a math or computer science degree to learn AI?
No. The roadmap starts from zero and teaches the math you actually need — linear algebra, calculus, and statistics — through the problems it solves, not as abstract theory.
What should I learn first for AI?
Start with what AI, machine learning, and deep learning actually are, then Python and working with data. Everything here is ordered by dependency, so you always know exactly what to learn next.
Can I learn AI on my own as a self-taught beginner?
Yes — this is built for self-taught learners: a clear ordered path, runnable code examples, and exercises, so you never have to guess what to study or in what order.
Do I need to know Python before starting?
No. Python is the first hands-on stage of the roadmap. If you already know it, you can skip ahead — every lesson lists its prerequisites.
Does it cover LLMs like ChatGPT?
Yes. The later stages cover how large language models work, prompt engineering, embeddings, RAG, fine-tuning, and building AI agents.