GenAI, agents and ML systems.

24 lessons (70.5 h of study) and 40 interview problems: RAG, agents, LLM serving, ML system design and implement-it coding.

16 to study, 2 to skim, 6 for later 6 easy, 18 medium, 16 hard

Roadmap

Read the modules in order. Each lesson has a worked example, runnable code, interview questions and a quiz.

Jump to problems

Foundations

6 lessons · 14 h

#1SkimBeginner 45 minHistory - how AI got to transformersAI did not jump straight to LLMs (large language models). Each generation of methods hit a wall, and the next idea was built to get past it: hand-written rules, then learning from data, then deep networks on GPUs, then attention, then transformers pretrained on huge amounts of text. This lesson walks that chain in order, so you can explain why a transformer looks the way it does and which problem each part solves.Open lesson
#2StudyBeginner 3.5 hNeural networks and backpropA neural network is a stack of simple units. Each unit multiplies its inputs by weights, adds a bias and applies a non-linear function. Training means finding weights that make a loss number small. Backpropagation computes how much each weight contributed to the loss, and gradient descent moves each weight a small step in the direction that lowers it. This lesson covers both, plus the basics interviewers check: activations, losses, optimisers, regularisation, normalisation and metrics.Open lesson
#3SkimBeginner 1 hRNNs, LSTMs, CNNsRNNs, LSTMs and CNNs are the three network designs that came before the transformer. An RNN reads a sequence one step at a time and carries a memory vector forward. An LSTM is an RNN with gates that protect that memory. A CNN slides small filters over an image. This lesson explains what each one solved, the limit each one hit, and which parts are still inside the models you use today.Open lesson
#4StudyBeginner 3 hLLM fundamentals - tokens, embeddings, decodingAn LLM (large language model) is a neural network trained to do one thing: predict the next token. This lesson follows a piece of text through the whole machine. Text becomes tokens, tokens become vectors, a stack of transformer blocks mixes those vectors, and a decoding rule picks the next token from the output probabilities. You will work each stage with small numbers.Open lesson
#5StudyIntermediate 2.5 hCoding simple attentionSelf-attention lets every token update its vector using the other tokens in the sequence. This lesson builds it in small steps: measure similarity with a dot product, turn the scores into weights with softmax, and take a weighted sum. Then it adds the three things a real layer needs: the scaling by the square root of d_k, the causal mask and multiple heads. Every step is worked with numbers you can check by hand.Open lesson
#6StudyBeginner 3 hHugging Face end to endHugging Face is where open models live and the set of Python libraries used to run them. This lesson follows one path from start to finish: find a model on the Hub, read its card and licence, load the tokenizer and the weights, format a chat with the model's template, generate text, measure the output, and save or push your own version. Every step names the object you get back.Open lesson

Build with LLMs

4 lessons · 13.5 h

Serving

1 lessons · 3.5 h

#11StudyIntermediate 3.5 hInference and serving - KV cache, batching, vLLMInference is running a trained model to produce output. Serving is doing that for many users at once, fast and cheaply. This lesson follows one request through an LLM server: the prefill and decode phases, the KV cache that makes decode fast, batching that keeps the GPU busy, quantization that shrinks the model, and the metrics and engines you use to run it in production.Open lesson

Agents

4 lessons · 15.5 h

Ship

3 lessons · 9 h

Later

6 lessons · 15.5 h

Interview problems

Each one has requirements, estimates, API, data model, design, diagram, deep dives, evaluation, failure modes and follow-ups.

GENAI

GenAI systems

· 22

RAG, agents, gateways, guardrails and evals: the questions asked for LLM application roles.

MLSD

ML system design

· 8

Serving, batching, vector search, recommendations and moderation at scale.

MLCODE

Implement it

· 10

Attention, tokenizers, sampling, LoRA and agent loops written from scratch in Python.

GenAI, agents and ML systems.

Roadmap

Foundations

Build with LLMs

Serving

Agents

Ship

Later

Interview problems

GenAI systems

ML system design

Implement it

Search

GenAI, agents and ML systems.

Roadmap

Foundations

Build with LLMs

Serving

Agents

Ship

Later

Interview problems

GenAI systems

ML system design

Implement it