WebAssembly AI Index

From portable bytecode to local intelligence.

WebAssembly AI Atlas is an independent crossover between systems programming and AI engineering. It follows one question through every layer: how can an AI capability be made portable, local, accelerated, isolated, and verifiable without hiding the bytes, memory, runtime, backend, or fallback that actually execute it?

Where this Atlas sits

• Low-Level Atlas I builds the machine model: C, memory, assembly, toolchains, operating systems, and concurrency foundations.
• Low-Level Atlas II builds applied systems from scratch: shells, allocators, parsers, Git, HTTP, and storage engines.
• Low-Level Atlas III covers sockets, protocols, large-scale concurrency, messaging, and long-lived networked systems.
• AI Atlas explains models, training, LLMs, RAG, agents, evaluation, safety, MLOps, product engineering, and general inference optimization.

This Atlas does not duplicate those foundations. It studies the execution layer where portable code, model artifacts, browser APIs, accelerators, and sandbox contracts meet.

Entry points

• Start Here — prerequisites, platform roles, and reading routes.
• Must Know — distinctions that prevent category errors.
• Phase 00: Orientation — establish the scope and verification contract.
• Reference Registry — track specifications, versions, support, and evidence.

Learning path

• Phase 01: WebAssembly Foundations — modules, toolchains, linear memory, ABI, and host interop.
• Phase 02: Inference from First Principles — tensor kernels, numerical correctness, SIMD, and threads.
• Phase 03: From Models to Browser Inference — formats, conversion, runtimes, and media pipelines.
• Phase 04: Hardware Acceleration — WebGPU, WebNN, and adaptive backend selection.
• Phase 05: Local-First AI — offline multimodal features, embeddings, and retrieval-first browser RAG.
• Phase 06: Beyond the Browser — WASI components and isolated AI tools.
• Always Active — performance, numerical verification, security, compatibility, and reproducibility.

Branch map

WebAssembly foundations

• WebAssembly Execution Model
• Toolchains & Language Targets
• Linear Memory, ABI & Host Interop

Inference from first principles

• Tensor Kernels & Inference from Scratch
• SIMD, Threads & Workers

Models to browser inference

• Model Formats, Conversion & Quantization
• Browser Inference Runtimes
• Preprocessing, Postprocessing & Media Pipelines

Hardware acceleration

• WebGPU for AI
• WebNN & Adaptive Backends

Local-first AI

• Local-First, Offline & Multimodal AI
• Browser LLMs, Embeddings & RAG

Beyond the browser

• WASI, Components & WASI-NN
• Sandboxed AI Tools & Plugins

Always active

• Wasm AI Performance, Security & Craftsmanship

WasmAI Workbench throughline

The Atlas converges on one local-first browser workbench. It progresses from a visible WAT module at v0, through a tiny inference engine and SIMD at v1–v2, real model formats and runtimes at v3, adaptive WebGPU/WebNN backends at v4, offline multimodal inference and semantic search at v5–v6, then portable WASI components and sandboxed tools at v7–v8. Every milestone must expose its backend and generate reproducible evidence.

How to navigate

Read phases in order when building the Workbench. Enter by branch when investigating a specific boundary, but follow its prerequisites and keep the always-active verification branch open from the first module onward. No page should require a claim of universal browser or hardware support.