Interactive ML and LLM Textbook
An interactive textbook for interview preparation: from linear regression and statistics to transformers, RAG and AI agents. Every topic is a plain-language explanation, an interactive visualization, interview questions and a mini-quiz.
The English version is in progress: the interview questions are fully translated, and topics are being translated one by one. Untranslated topics link to the Russian original.
Machine Learning Fundamentals · interview questions
The foundation: how a model learns from data, linear and logistic regression, gradient descent, overfitting and regularization, metrics, trees and ensembles, clustering and dimensionality reduction.
- What Is Machine Learning (in Russian)
- Types of Learning: Supervised, Unsupervised and Reinforcement (in Russian)
- Linear Regression (in Russian)
- Gradient Descent (in Russian)
- Logistic Regression and Classification (in Russian)
- Bayes' Theorem and Naive Bayes (in Russian)
- Overfitting and Regularization (in Russian)
- Bias–Variance: Decomposing the Error (in Russian)
- Quality Metrics (in Russian)
- Cross-Validation and Hyperparameter Tuning (in Russian)
- Decision Trees and Ensembles (in Russian)
- SVM and the Kernel Trick (in Russian)
- kNN and Clustering (in Russian)
- The EM Algorithm and Gaussian Mixtures (GMM) (in Russian)
- Dimensionality Reduction and PCA (in Russian)
- Features and Data Leakage (in Russian)
- Time Series and Forecasting (in Russian)
- Reinforcement Learning: Q-learning and PPO (in Russian)
Statistics for ML
The statistics you need to compare models honestly: samples and the central limit theorem, standard error, confidence intervals with a calculator, p-values and hypothesis testing.
- Sampling, the CLT and Standard Error
- Confidence Intervals (in Russian)
- P-Values and Hypothesis Testing (in Russian)
Neural Networks · interview questions
Neural networks from a single neuron to convolutional and recurrent architectures: activation functions, backpropagation, optimizers, batches, dropout and batchnorm.
- From a Neuron to a Multilayer Perceptron (in Russian)
- Activation Functions (in Russian)
- Backpropagation (in Russian)
- Optimizers: SGD, Momentum, Adam (in Russian)
- Training in Practice: Batches, Dropout, Batchnorm (in Russian)
- Recognizing Digits from Pixels (in Russian)
- Convolutional Neural Networks (CNN) (in Russian)
- Recurrent Neural Networks (RNN, LSTM) (in Russian)
Language Models: From N-grams to the Transformer · interview questions
How models work with text: n-grams, tokenization and BPE, embeddings and word2vec, the attention mechanism, the transformer architecture and modern building blocks RoPE, GQA and MoE.
- Bigrams and N-grams (in Russian)
- Tokenization and BPE (in Russian)
- Embeddings and word2vec (in Russian)
- The Attention Mechanism (in Russian)
- The Transformer Architecture (in Russian)
- Modern Architectures: RoPE, MoE, GQA (in Russian)
LLMs: Training and Inference · interview questions
The life cycle of a large language model: pretraining and scaling laws, fine-tuning and LoRA, RLHF and DPO, reasoning models, sampling, inference, distillation, evaluation and working with context.
- Pretraining and Scaling Laws (in Russian)
- Fine-tuning: SFT, LoRA, PEFT (in Russian)
- Alignment: RLHF and DPO (in Russian)
- Reasoning Models and Test-Time Compute (in Russian)
- Sampling: Temperature, Top-k, Top-p (in Russian)
- Inference: KV-Cache, Quantization, Speculative Decoding (in Russian)
- Distillation and Model Compression (in Russian)
- Evaluating LLMs: Benchmarks and LLM-as-Judge (in Russian)
- Working with Context and Message Markup (in Russian)
- Prompt Patterns and Structured Outputs (in Russian)
Multimodal Models · interview questions
Models beyond text: how LLMs see images and hear speech, omni models, OCR, and image and video generation from VAEs and GANs to diffusion models.
- How LLMs See Images (in Russian)
- Audio: Speech Recognition and Synthesis (in Russian)
- Omni Models: All Modalities in One (in Russian)
- OCR: From Classic Methods to Recognition with LLMs (in Russian)
- VAEs and GANs: Generation Before Diffusion (in Russian)
- Diffusion Models: How Stable Diffusion Draws (in Russian)
- Video Models: Generation with Temporal Consistency (in Russian)
Hardware and Deployment
Hardware and deployment: GPUs and VRAM estimates, distributed training with DP, TP, PP and ZeRO, choosing between local models, the cloud and APIs.
- GPUs and VRAM: Hardware for Neural Networks (in Russian)
- Distributed Training: DP, TP, PP and ZeRO (in Russian)
- Local Models vs SaaS and APIs (in Russian)
Agents and Tools · interview questions
AI agents in practice: tool calling, the agent loop and stopping criteria, memory and planning, multi-agent systems, the MCP protocol and agent evaluation.
- Tool Calling: How a Model Calls Tools (in Russian)
- The Agent Loop and Stopping Criteria (in Russian)
- Memory, Planning and Multi-Agent Systems
- MCP: The Standard for Connecting Tools (in Russian)
- Evaluating Agents: pass@k and Trajectories (in Russian)
RAG: Retrieval + Generation · interview questions
RAG from idea to production: vector search and ANN indexes, vector databases, chunking, hybrid search and reranking, MMR, training embedding models and quality evaluation.
- What RAG Is and Why You Need It (in Russian)
- Vector Search and ANN Indexes (in Russian)
- Vector Databases and Their Types (in Russian)
- Chunking and Data Preparation (in Russian)
- Hybrid Search and Reranking (in Russian)
- MMR and Result Diversity (in Russian)
- Embedding Models: How They Are Trained (in Russian)
- Evaluating RAG Quality (in Russian)
Production and Quality · interview questions
ML in production: ML system design in interviews, recommender systems and ranking, monitoring and drift, A/B tests, safety and guardrails.
- ML System Design in Interviews (in Russian)
- Recommender Systems (in Russian)
- Ranking: Learning to Rank (in Russian)
- Monitoring and Data Drift (in Russian)
- A/B Tests and Product Metrics (in Russian)
- Safety: Guardrails and Prompt Injection (in Russian)