Skip to main content
Version: Next 🚧

Roadmap

Murali's direction is Python-first authoring for deterministic technical storytelling, backed by a Rust runtime for rendering performance. It should visualize recorded model and system state rather than become an inference framework, and it should preserve explicit author control while making common educational compositions easier to build.

This page focuses on future work. Released capabilities are documented in the guides and release notes.

Python Stability Toward 0.5.0​

Murali 0.3.0 established the current Python-first package shape: murali-engine as the Python frontend over the Rust runtime, and murali-kit for themes, named colors, examples, and teaching views. The next milestone is turning that early surface into something stable enough to keep.

  • Keep constructor names, validation, return values, and module registration consistent across the Python frontend.
  • Fill Python gaps when they block real examples, documentation, or murali-kit teaching views.
  • Use kit examples as the backbone for public documentation and API checks.
  • Keep the Rust crate focused on the engine while preserving enough extension points for future authoring layers.
  • Treat breaking Python cleanup as acceptable before 0.5.0, then move toward compatibility.
  • Add release checks for Python import, Python examples, murali-kit examples, Maturin wheel builds, docs samples, and normal Rust tests.

Reliability And Quality Gates​

  • Cross-platform and headless wgpu runtime coverage.
  • Golden-image tests for core rendering and representative AI scenes.
  • Property tests for mesh limits, transformed bounds, capture ordering, and invalid input.
  • Structured validation across remaining token, attention, label, color, dimension, and route APIs.
  • A warning-free library build, broader strict Clippy, dependency auditing, and validated docs samples.

AI Teaching Semantics​

  • A domain RFC defining snapshots, tokens, operations, semantic IDs, trace events, and data ownership.
  • General image and image-grid compositions for datasets, feature maps, and multimodal lessons.
  • Residual-stream, MLP, and mixture-of-experts semantics.
  • An autoregressive-generation composition connecting successive next-token distributions, selected tokens, appended context, and the repeat-until-stop streaming cycle.
  • Computation graphs with deterministic forward and reverse-mode playback.
  • Parameter, gradient, loss, optimizer, activation, feature-map, and attribution views.
  • A complete semantic backpropagation lesson spanning equations, tensors, networks, and plots.
  • Recorded agent and RAG events for retrieval, tools, memory, retries, branches, and handoffs.
  • Aggregation, sampling, clipping, and virtualization for long sequences and dense models.

Narration And Production​

  • First-class narration segments, audio tracks, cue bookmarks, captions, and subtitles.
  • Cue-aligned clips and machine-readable render manifests.
  • Reusable teaching layouts for title, equation, diagram, comparison, and recap shots.
  • A deterministic author-render-inspect-revise workflow using screenshots and scene metadata.

Authoring Architecture​

  • Stable semantic scene names alongside numeric tattva IDs.
  • A versioned JSON SceneSpec with source locations, validation, and repair diagnostics.
  • murali validate, render, doctor, and inspect commands built on that declarative boundary.
  • Tighter public module boundaries and smaller animation modules with clear ownership.
  • Scene-owned themes and deliberate removal or integration of unused global state.
  • Consolidated process-diagram APIs and overridable lesson templates.

Renderer And Scale​

  • Order-independent transparency for intersecting translucent 3D geometry.
  • Reused line buffers and bind groups, plus batched uniform uploads.
  • A deliberate renderer mesh-cache policy and complete text-resource caching.
  • Tiling and resource policies for large raster-backed text and code surfaces.
  • Separate CPU projection and GPU upload measurements, backed by representative AI benchmarks.

Project Sustainability​

  • A deliberate crates.io examples-packaging policy.
  • Contribution and security policies, changelog discipline, and release automation.
  • Continued synchronization between the website, API docs, examples, and release notes.

The detailed maintainers' version of this plan lives in the repository's ROADMAP.md.