AI Visualization

Murali is a Python animation engine for AI education and technical visual storytelling. Teaching views such as attention matrices and neural diagrams live in Murali Kit. The runtime is Rust; day-to-day authoring is Python.

The long-term target is to grow that visualization surface steadily through the end of 2030.

Murali is not trying to become an inference framework. It is meant to visualize authored or recorded state: tensors, tokens, matrices, distributions, model traces, optimization steps, geometric intuition, and the surrounding mathematical structure.

Current Status

• AI visualization exists today. Murali includes components for attention matrices, context windows, KV cache, next-token distributions, neural-network diagrams, transformer blocks, normalization views, tensor snapshots, semantic tensor transitions, decision boundaries, and AI traces.
• Linear algebra is experimental. The experimental feature exposes reusable linear-algebra pieces for vectors, transformed grids, matrix-vector flow, column combinations, projections, orthogonality, dimension badges, transform composition, and basis changes.
• Other education categories are planned. Probability, statistics, calculus, optimization, information theory, data geometry, and systems/agentic AI will be expanded over the next several years as reusable teaching components prove themselves in real examples.

Category Surface

Category	Functional anchor today	Direction
Basic math	equations, matrices, labels, LaTeX, Typst, axes	algebra steps, intervals, symbolic transforms
Linear algebra	experimental linear_algebra components	rank, eigenspaces, SVD, subspaces
Probability	NextTokenDistribution, graphs, tables	distributions, sampling, Bayes diagrams
Statistics	NormalizationView, DecisionBoundaryPlot, scatter plots	regression, uncertainty, metrics
Calculus	function graphs, parametric curves/surfaces, vector fields	derivatives, integrals, gradients
Optimization	surfaces, vector fields, traced paths	descent, optimizers, loss landscapes
Information theory	next-token probabilities, tensors, equations	entropy, KL, cross-entropy, calibration
ML components	tensors, decision boundaries, signal flow	datasets, embeddings, losses, evaluations
Deep learning	neural-network, transformer, attention, normalization views	gradients, residuals, feature maps
Transformers and LLMs	context, tokens, attention, KV cache, sampling, traces	generation loops, RAG, tools, memory
Data geometry	scatter plots, surfaces, tensor projections	embeddings, PCA, clustering, manifolds
Systems and agentic AI	agentic flow charts, traces, context windows	tools, retries, branches, evaluations

Import Surface

Murali's collection is the author-facing collection of reusable visual tattvas. Its implementation families remain organized by primitive family: ai, maths, text, layout, storytelling, and related modules. Alongside those, Murali exposes domain-oriented category modules directly under collection:

use murali::collection::ai::transformers_llms::*;
use murali::collection::maths::probability::*;
use murali::collection::maths::data_geometry::scatter_plot::ScatterPlot;

These category modules are re-export layers over real components. They are meant for lesson authors who think in teaching domains rather than implementation folders. A component still has one implementation owner; a domain facade is only a discovery surface. The experimental linear-algebra module is available only when the experimental feature is enabled.

Support pieces stay in implementation families such as primitives, text, maths, composite, layout, storytelling, table, and utility. Domain facades are kept only where they clarify the teaching subject.

Source Architecture

The source-visible category READMEs live in src/frontend/collection. They document the category ownership rules and intended 2026-2030 evolution without pretending that every planned area is already stable.

The rule for this surface is simple: add reusable teaching primitives and composites, not one-off scene shortcuts. A category can stay planned until a real lesson proves that a component belongs in Murali.