AI Interview Prep
A visual study path for AI/ML interviews — 80 animated diagrams across 8 layers, from the fundamentals to LLMs, systems, and security.
A complete visual curriculum for AI/ML interviews — 80 animated diagrams across 8 layers. Start at the top with the fundamentals, work down through the algorithms and architectures, and tap any tile to open it full-size. Each one is a single detailed picture you can study the night before.
1 · Foundations
The core ideas every ML interview opens with.
The Machine Learning LandscapeSupervised, unsupervised, RL — and which to use.
The ML Project LifecycleFrom problem framing to a production loop.
Bias–Variance TradeoffUnderfitting, overfitting, and the sweet spot.
Train / Validation / Test & CVSplit data so your scores stay honest.
Gradient DescentHow models actually learn — stepping downhill.
Loss FunctionsWhat you minimize: MSE, MAE, cross-entropy.
Classification MetricsConfusion matrix, precision, recall, F1, ROC-AUC.
RegularizationL1, L2, dropout — fighting overfitting.
2 · Supervised Learning
The classic algorithms you must be able to explain.
Linear RegressionPredict a number with the best-fit line.
Logistic RegressionPredict a probability with the sigmoid.
Decision TreesIf/else splits chosen by information gain.
Random Forests & BaggingMany trees vote — variance down.
Gradient Boosting (XGBoost)Sequential trees that fix prior errors.
Support Vector MachinesThe widest-margin separating boundary.
KNN & Naive BayesTwo simple, strong baselines.
3 · Unsupervised Learning
Finding structure when there are no labels.
K-Means ClusteringGroup unlabeled data into k clusters.
PCA / Dimensionality ReductionSqueeze features, keep the variance.
4 · Deep Learning
From a single neuron to the transformer block.
The Neuron & PerceptronThe building block of every neural net.
BackpropagationHow gradients flow back to every weight.
Activation FunctionsWhy nonlinearity gives nets their power.
Convolutional Neural NetworksFilters that learn a feature hierarchy.
RNNs & LSTMsMemory across a sequence.
The Transformer BlockThe architecture behind every LLM.
How Attention WorksQ, K, V — the heart of the transformer.
5 · Transformers & LLMs
How modern language models are built and trained.
How Tokenization WorksText becomes numbers (BPE).
Embeddings & Vector SearchMeaning as vectors you can search.
What Happens When You Ask ChatGPTA prompt's full round trip.
LLM Inference & ServingHow a token actually gets generated.
Mixture-of-Experts (MoE)Scale params, not compute.
LLM Fine-tuning PipelineAdapting a base model to your task.
How RLHF Aligns a ModelTeaching a model what we prefer.
RAG vs Fine-tune vs PromptWhich adaptation strategy to pick.
6 · RAG, Agents & Generative
Retrieval, tool-use, agents, and image generation.
Agentic RAG PlatformPlan, search, call tools, ground the answer.
Knowledge Graph / GraphRAGRetrieval that understands relationships.
MCP & Tool-CallingHow agents reach tools and data.
Multi-Agent OrchestrationA supervisor routes specialist agents.
How an AI Agent Thinks (ReAct)Reason, act, observe, repeat.
AI Agent MemoryHow an agent remembers.
How Diffusion Models Create ImagesFrom noise to an image.
Voice AI PipelineSpeech → LLM → speech, in under a second.
7 · ML Systems & System Design
The infra and design-interview questions.
Modern Data PlatformIngest → lake → transform → govern → serve.
Lakehouse ArchitectureWarehouse + lake, unified.
Streaming Data PlatformData in motion, not at rest.
Feature StoreOne source of truth for features.
Realtime ML / Fraud DetectionStream → features → score in ms.
Recommendation SystemHow the feed gets picked.
AI Gateway / LLM RouterOne front door for every model.
MLOps CI/CDPush → eval gates → canary → promote.
AI Evals & ObservabilityLog, score, trace the why, catch drift.
Experimentation / A-B TestingDid the change actually help?
Model Compression & DistillationBig model into small & fast.
GPU Supercluster AnatomyWhere models are trained.
Inside a Frontier Training RunPretrain → eval → post-train → release.
8 · AI Security & Governance
Securing AI systems — the bonus track that sets you apart.
Securing the LLM App StackOWASP Top 10 for LLMs, every layer.
Prompt Injection DefenseStopping the #1 LLM attack.
LLM Data Privacy & PIIKeep secrets out of the model.
Confidential AIInference nobody can peek into.
AI/ML Supply Chain SecurityTrust what you didn't train.
Zero Trust for AI AgentsIdentity, policy, least privilege.
AI Security Posture (AI-SPM)Secure your whole AI estate.
Shadow AI / DLPDiscover AI tools, redact secrets.
AI Email SecurityStopping phishing with AI.
Adversarial ML DefenseAttacks on the model itself.
Deepfake DetectionIs this real?
AI-Powered Threat IntelFrom noise to named threats.
AI Red Teaming PlatformBreak models before attackers do.
AI-Powered SOCML triage mapped to MITRE ATT&CK.
AI Governance & ComplianceShip AI you can defend.
9 · Landmark Architectures
Seminal papers drawn as architectures — the deep-dives behind the concepts.
Generative Adversarial NetworksTwo networks locked in a duel.
Variational Autoencoder (VAE)A smooth, samplable latent space.
StyleGANControl image style at every scale.
pix2pix — Image-to-ImageA conditional GAN for paired images.
ResNet — Residual LearningSkip connections enable very deep nets.
Vision Transformer (ViT)An image is worth 16×16 words.
DETR — Detection TransformerObject detection as set prediction.
LoRAFine-tune with tiny low-rank adapters.
QLoRAFine-tune a huge model on one GPU.
10 · Adversarial Attacks
How models get fooled — the attack papers, visualized.
FGSMOne gradient step fools a classifier.
PGD AttackThe strongest first-order attack.
Black-Box AttacksFool a model with only its labels.
Follow along on LinkedIn.