Learn AI the way engineers learn: by building
A hands-on path from data analysis and classical machine learning (regression, classification, clustering, forecasting) through neural networks, transformers, LLMs and RAG to MLOps and ML system design, written for software developers, whatever language or framework you use. Every lesson connects the idea to real web, mobile, data and cloud products, then gives you something to build.
What every lesson includes
Read the idea, see it as a diagram, try it with real inputs, build it, then test yourself.
Architecture diagram
How data flows through the model or system, step by step.
Real-world examples
Every idea explained with an everyday example first, then real inputs and outputs.
Coding exercises
Starter code with hints, mostly in Python, plus web, mobile and SQL exercises.
A mini project
A small build with an implementation checklist you tick off.
A quiz
Checked on our server, with an explanation for every answer.
Spaced review
Flashcards that come back just before you would forget them.
Research summaries
The papers behind each idea, in plain English, with links.
Your notes
Private notes filed by lesson, searchable from one place.
The curriculum
About 39 hours of reading, plus the exercises and projects. Take the tracks in order, or start with the lesson your current project needs.
Foundations
What a model is, how it learns from data, how training differs from inference, and how to pick the right kind of model for a problem before writing any code.
How AI models actually work
AI, machine learning and deep learning explained as code: a model is a function with millions of tunable numbers, learned from examples instead of written by hand.
Training vs inference
Why training is a slow, expensive, GPU-heavy batch job and inference is a fast, repeated request, and what that means for cost, latency, hardware and the way you design features.
Datasets, labels and features
Where training data comes from, how to label and split it, what features and leakage are, and why most model quality problems are really data problems.
Neural networks from first principles
Neurons, layers, activation functions, backpropagation and the training tricks that make deep networks work, built up from a single weighted sum to a working network in PyTorch.
When to use each type of model
A decision guide that maps problems to approaches (rules, classical ML, deep learning, pretrained models, LLM prompts, RAG or fine-tuning) with the trade-offs in cost, data and risk.
Maths and statistics for ML
The maths every model runs on, taught through code: vectors and matrices, gradients and optimisers, probability, uncertainty and hypothesis tests.
Linear algebra for machine learning
Vectors, matrices, dot products, norms, matrix multiplication, eigenvectors and SVD: the maths every model runs on, explained through NumPy code and the data you already store.
Calculus and optimisation for ML
Derivatives, gradients, the chain rule, loss functions and gradient descent with momentum and Adam: how a model actually finds good weights, and how to tell when it will not.
Probability and statistics for ML
Distributions, expectation and variance, Bayes' theorem, maximum likelihood, sampling, confidence intervals, the bootstrap and hypothesis tests: the reasoning behind every metric and every A/B test.
Data: collection, analysis and processing
Where data comes from and how to turn it into something a model can learn from: collection and labelling, exploratory analysis, cleaning, feature engineering and pipelines that scale.
Data collection, sampling and labelling
Where ML data comes from: product instrumentation, logs, APIs, public datasets and scraping; sampling bias; labelling at scale with agreement checks, active learning and weak supervision; and consent.
Exploratory data analysis
Profiling, distributions, outliers, correlations, segment and target analysis, time trends and Simpson's paradox with pandas and SQL: understanding a dataset before any model sees it.
Data cleaning and preprocessing
Missing values and imputation, outliers, duplicates, type fixes, categorical encoding, scaling and transforms, all wrapped in scikit-learn pipelines so nothing leaks between training and test data.
Feature engineering and selection
Aggregates, windows, ratios, time, text and geographic features; interactions and binning; filter, wrapper and embedded selection; permutation importance; and point-in-time correct feature stores.
Data pipelines and processing at scale
ETL and ELT, batch and streaming, warehouses and lakes, Parquet, pandas, Polars, DuckDB and Spark, orchestration, idempotent backfills, data validation and dataset versioning.
Classical ML: regression and classification
The supervised algorithms behind most production ML, with how each one works and when to use it: validation and tuning, linear and logistic regression, k-NN, Naive Bayes, SVMs, decision trees, random forests and gradient boosting.
The ML workflow, validation and tuning
Baselines, the bias-variance trade-off, learning curves, k-fold, stratified, group and time-series cross-validation, nested CV, grid, random and Bayesian tuning, and error analysis.
Linear regression and regularisation
Least squares, the normal equation, assumptions and residual diagnostics, interpreting coefficients and log transforms, polynomial features, and Ridge, Lasso and ElasticNet regularisation.
Regression algorithms: GLMs, quantiles and intervals
The regression toolbox beyond OLS: Poisson, Gamma and Tweedie GLMs, robust regression, k-NN, SVR, Gaussian processes and tree ensembles, quantile regression, conformal prediction intervals and extrapolation.
Logistic regression and classification
The sigmoid, log-odds and cross-entropy, odds ratios, multiclass with softmax, decision thresholds chosen from costs, probability calibration, and handling imbalanced classes.
k-nearest neighbours and Naive Bayes
Two simple, surprisingly strong classifiers: k-NN with distance metrics, scaling, the curse of dimensionality and fast neighbour search; and Naive Bayes for text with smoothing and log probabilities.
Support vector machines and kernels
Maximum-margin classifiers, support vectors, soft margins and the hinge loss, the kernel trick with RBF and polynomial kernels, tuning C and gamma, SVR and one-class SVMs, and when SVMs still win.
Decision trees
How CART trees split with Gini, entropy and variance reduction, controlling depth and pruning, interpretability and instability, extrapolation limits, and exporting a tree as code.
Ensembles: random forests and gradient boosting
Why averaging models works, bagging and random forests with out-of-bag error, AdaBoost and gradient boosting, XGBoost, LightGBM and CatBoost, early stopping, monotonic constraints, tuning and stacking.
Unsupervised and specialised ML
Learning without labels and the specialised problems every product meets: clustering, dimensionality reduction, anomaly detection, time-series forecasting and recommender systems.
Clustering and segmentation
Unsupervised grouping with k-means and k-means++, hierarchical clustering and dendrograms, DBSCAN and HDBSCAN, Gaussian mixtures, choosing k, evaluating clusters and turning them into useful segments.
Dimensionality reduction
PCA and explained variance, truncated SVD for text, LDA and NMF, t-SNE and UMAP for visualisation and how not to misread them, autoencoders, and shrinking embeddings for faster search.
Anomaly detection
Point, contextual and collective anomalies; robust statistical baselines and seasonal comparisons; Isolation Forest, LOF, one-class SVM and reconstruction error; evaluation with few labels; and alerting without fatigue.
Time-series forecasting
Trend, seasonality and autocorrelation; naive and seasonal-naive baselines; exponential smoothing, ARIMA and Prophet; gradient boosting with lag features across many series; backtesting, MASE and WAPE; probabilistic and hierarchical forecasts.
Recommender systems
Explicit and implicit feedback, popularity baselines, content-based and collaborative filtering, matrix factorisation with ALS, two-tower retrieval and ranking, cold start, recall@k and NDCG, and feedback loops.
Core architectures
The building blocks behind modern AI: embeddings, convolutional networks for images, recurrent networks for sequences, transformers and attention, and reinforcement learning.
Embeddings and vector search
How models turn text and images into vectors where distance means similarity, how to search millions of them quickly, and how to add semantic search to an app and its database.
Convolutional neural networks (CNNs)
How convolutions, pooling and feature maps let networks see, why pretrained CNNs make image features cheap to build, and how to ship one to a phone.
Recurrent networks: RNNs, LSTMs and GRUs
How recurrent networks carry state through a sequence, why plain RNNs forget, how LSTM and GRU gates fix that, and where these models still make sense after transformers.
Transformers and attention
Tokens, embeddings, self-attention, multi-head attention and the transformer block, explained step by step, plus encoder, decoder and encoder-decoder variants and why they scale so well.
Reinforcement learning
Agents, environments, rewards and policies; Q-learning and policy gradients; and the practical forms you will actually meet: bandits for product decisions and learning from human feedback for LLMs.
Modern AI models
Large language models, multimodal models, retrieval-augmented generation, fine-tuning and evaluation: how today's models work and how to make them reliable for your own data.
Large language models
How LLMs are pretrained, instruction-tuned and run; what tokens, context windows and sampling mean in practice; and how to call them reliably from your application with structured output and streaming.
Multimodal models
How models combine text with images, audio and video: shared embedding spaces such as CLIP, vision-language models that read screenshots and documents, speech models such as Whisper, and image generation with diffusion.
Retrieval-augmented generation (RAG)
Ground LLM answers in your own documents: chunking, embedding, retrieval, re-ranking, prompt assembly and citations, plus how to measure retrieval and answer quality separately.
Fine-tuning and adaptation
When fine-tuning is worth it, how supervised fine-tuning, LoRA and QLoRA work, how to prepare training data, and where preference tuning and distillation fit, compared with prompting and RAG.
Evaluating models and LLM apps
Metrics for classifiers, regressors and retrieval, plus how to evaluate LLM features with test sets, assertions, LLM-as-judge and human review, so every prompt or model change is measured instead of guessed.
Shipping AI
Serving models in production, building agents that call your APIs, and the patterns, guardrails and costs of AI features in real web, mobile and backend products.
Deploying and serving models
From notebook to production: packaging, serving APIs, batching, quantisation, GPUs versus CPUs, caching, monitoring, versioning and cost control, with patterns for background queues, Python services and mobile.
Agents and tool use
How LLMs call functions you define, how agent loops plan and act, and how to give an LLM safe access to your app's data and actions with schemas, permissions, confirmations and logging.
AI in production software: patterns and use cases
The real-world patterns for AI features in web, mobile, backend and cloud products: where AI fits, UX for uncertainty, privacy and security, cost control, team workflow and a launch checklist.
Elite ML engineering
The skills that separate senior ML engineers: training deep networks well, scaling across GPUs, interpretability and fairness, causal inference, MLOps and end-to-end ML system design, finishing with a capstone project.
Training deep networks well
A reliable training recipe: overfit one batch first, initialisation and normalisation, AdamW with warm-up and cosine schedules, regularisation and augmentation, transfer learning, mixed precision and systematic debugging.
Scaling training and inference
GPU memory and compute arithmetic, data parallelism with all-reduce, ZeRO and FSDP sharding, tensor and pipeline parallelism, activation checkpointing, scaling laws and compute-optimal training, and efficient large-model inference.
Interpretability, fairness and responsible ML
Global and local explanations with permutation importance, partial dependence, SHAP, LIME and counterfactuals; interpretable models; fairness metrics and their trade-offs; mitigation, model cards and human oversight.
Causal inference and advanced experimentation
Correlation vs causation, potential outcomes and causal graphs, confounders and colliders, rigorous A/B tests with SRM checks and CUPED, propensity scores, difference-in-differences, and uplift modelling.
MLOps: the model lifecycle
Experiment tracking and reproducibility, model registries and promotion, testing ML systems, CI/CD and continuous training, drift metrics such as PSI, monitoring with delayed labels, retraining strategies and incident response.
ML system design (capstone)
A framework for designing ML systems end to end: requirements, framing, metrics, data, features, models, serving, evaluation and monitoring, with fraud detection, ranking and LLM-assistant case studies and a capstone project.
Easy to continue every week
Set a weekly goal of one to seven lessons. The workspace remembers your current lesson, counts your week streak, and keeps a short review queue so earlier topics stay fresh. Fifteen minutes on a busy day still counts.
This week
Your weekly goal, streak, current lesson and what is due.
Roadmap
Four tracks, every lesson and where you are with each.
Model comparison
Architectures, adaptation methods, serving options, model families and metrics side by side.
Completed topics
What you have finished, with your best quiz scores.
About your account
An account is free and only needed for Learn AI; every tool on ElysiumHub still works without one. We store your name, email address, password (hashed) and what you do in the workspace: lesson progress, quiz answers, practice tasks and links, notes and your review schedule. Your notes are visible only to you. You can delete your account and all of this at any time from My account. Details are in the privacy policy.
