Interactive AI engineering diagrams

26 animated, interactive figures explaining how AI systems actually work — attention, tokenization, embeddings, the agent loop, RAG, backpropagation and more. Each has its own page.

Agents

• The Agent Loop: Observe → Think → Act — Animated ReAct loop showing the observe-think-act cycle with tool calls and responses.
• Multi-Agent Orchestration Patterns — Orchestration patterns: sequential, parallel, supervisor, and hierarchical agent topologies.

RAG & Embeddings

• RAG Pipeline: Ingest → Retrieve → Generate — End-to-end RAG pipeline with document chunking, embedding, vector store, retrieval, and generation.
• Embedding Space Projection — 2D embedding space with semantic clusters, showing king − man + woman = queen arithmetic.

Transformers

• Self-Attention Matrix — N×N attention grid lighting up token-to-token relationships with softmax weighting.
• BPE Tokenizer: Merge Rules in Action — Byte-pair encoding animation showing character-level merging into subword tokens.
• Transformer Block Data Flow — Data flowing through residual connections, multi-head attention, and feed-forward layers.
• KV Cache Growth During Generation — Key-value cache expanding left-to-right as tokens are generated autoregressively.

Training

• Training Loss Curve — Animated loss curve descending over training steps with learning rate annotations.
• Gradient Descent: Rolling Down the Loss Surface — A ball rolling down a parabolic loss surface, following the negative gradient at each step until it converges to the minimum.
• Backpropagation: Error Signals Flow Backward — Forward pass activating neurons left-to-right, then gradient signals flowing backward through the network, updating each weight by its contribution to the loss.

Tools

• MCP Server Architecture — Model Context Protocol server with tools, resources, and prompts exposed to client.

Evaluation

• AI Evaluation Pipeline — Evaluation pipeline from test cases through scoring, regression detection, and dashboards.

Security

• Defense-in-Depth Guardrail Stack — Layered guardrail architecture: input filter, model safety, output validation, monitoring.

Computing

• Bits to Bytes to a Number — Eight individual bits flipping on and off, grouping into one byte, then resolving to the decimal number they encode.
• CPU vs. GPU: Few Powerful Workers vs. Thousands of Simple Ones — A handful of CPU cores processing a large grid sequentially, racing against thousands of small GPU cores processing it all at once.
• A Variable Is a Name Tag, Not a Box — A name tag "x" detaching from one value and re-attaching to a newly computed value, showing assignment as rebinding rather than mutation.

Mathematics

• Multiplying by a Negative Is a Rotation — Points on a number line swinging 180° about zero when multiplied by -1, then stretching further out when multiplied by -2.
• A Function Never Has to Think Twice — An animated machine turning different inputs into single, predictable outputs, then jamming when one input is asked to produce two different outputs.
• The Logarithm Is the Exponential, Reflected — Paired exponential and logarithm curves mirrored across y = x, reshaping together as the base changes.
• Sine and Cosine, Unrolled from a Circle — A point traveling around the unit circle with its cosine and sine projections shown live, while the sine value simultaneously unrolls into a wave trace.
• A Secant Line Converging Into a Tangent — A secant line through two points on a curve pivoting into the tangent line as the second point slides toward the first, with the live slope value converging numerically.

Data

• Split, Apply, Combine — Hadley Wickham's split-apply-combine framing animated literally: a table splitting into colored groups, each group collapsing into one aggregated value, then recombining into a final result.
• Broadcasting: Stretched, No Copy Made — A (4,) array visually stretching into a ghosted (3, 4) shape to align with a matching array, illustrating that NumPy broadcasting reuses values in place rather than copying them.
• Start From the Shape of Your Data — An animated decision guide where a data shape (one numeric, two numeric, categorical+numeric, part-to-whole) lights up a path to its recommended chart types.

Deployment

• Inference Latency: Prefill vs. Decode — Token-by-token generation timeline showing the first token paying the prefill cost while subsequent tokens decode in a fraction of the time.

Create a free account — the opening phases of 24 of 30 courses are free, no credit card. Or see Pro pricing.

All courses · Pricing · About · FAQ · Glossary