Making the web accessible for every mind.
AI-powered Chrome extension that makes any web article accessible — on-device LLM via WebGPU, zero cloud, zero backend.
AMD Slingshot 2026 Submission — Theme: AI for Social Good Category: Assistive tools for vision, speech, reading, and neurodiverse learners.
Elu (short for Elucidate) is a Chrome extension that transforms any web article into an accessible reading experience — on-device, offline, and privacy-preserving. It runs a quantized large language model directly in the browser using WebLLM and WebGPU, eliminating the need for cloud APIs or user data transmission.
Built for a wide spectrum of cognitive and sensory needs: dyslexia, ADHD, low vision, language learners, and anyone who benefits from a calmer, clearer reading environment. Every feature — AI simplification, smart chunking, focus mode, bionic reading, text-to-speech, and colour themes — works entirely on your device. Nothing is ever sent to a server.
- The Problem
- Solution Overview
- Architecture
- Use Cases
- Feature Reference
- Accessibility Design Principles
- Technology Stack
- AMD Hardware Integration
- Getting Started
- Keyboard Shortcuts
- Privacy Guarantee
- Contributing
- License
Approximately 1 in 5 people has a language-based learning difference such as dyslexia. An estimated 366 million people worldwide live with ADHD. Countless others navigate the web with low vision, cognitive fatigue, or as non-native speakers. Yet the modern web is built with one reader in mind: a neurotypical adult with perfect vision.
Dense paragraphs, distracting sidebars, autoplay videos, and complex vocabulary create invisible barriers for a significant fraction of the population. Existing assistive tools are either siloed, superficial, or rely on cloud processing that creates privacy risks.
Elu intercepts the reading experience at the browser level and provides a layered set of interventions:
| Layer | What Elu Does |
|---|---|
| Language | Rewrites content to 5 configurable reading levels using an on-device LLM |
| Cognition | Breaks articles into 150-word chunks with progress tracking and bookmarks |
| Focus | Dims distractions, blocks animations, and provides ambient sound assistance |
| Vision | Applies 13 high-contrast colour themes, spacing controls, and OpenDyslexic font |
| Perception | Bionic reading (bold-first-half rendering) to guide the eye efficiently |
| Audio | Paragraph-by-paragraph text-to-speech with variable speed and voice selection |
| Comprehension | Inline glossary via double-click word lookup |
All processing occurs locally. No page content ever leaves the browser.
- Offscreen Document + Web Worker: The WebLLM engine runs in a dedicated Web Worker inside Chrome's
offscreendocument, keeping WebGPU inference off the UI thread entirely and preventing extension page freezes. - Background service-worker as router: All LLM requests from the content scripts are routed through the background service worker, which manages offscreen document lifecycle, model selection, and exponential-backoff retries.
- Mozilla Readability for content extraction: A shared
content-extractor.jsmodule uses@mozilla/readabilitywith a CSS-selector fallback to reliably locate the main article body, independent of page structure. - System prompt library: Three prompt families (
textClarity,focusStructure,wordPattern) each with five intensity levels (1–5) provide 15 distinct rewrite personalities delivered fromprompts.jswithout duplicating logic in the content layer. chrome.storage.syncfor settings: All user preferences are synced across devices at the storage layer; no settings server is required.
Elu serves four primary reader profiles, each mapped to a tailored combination of features:
| Reader | Primary Needs | Key Features |
|---|---|---|
| Dyslexia | Letter confusion, reading fatigue | Bionic Reading, OpenDyslexic font, Word Pattern simplification, Cream Paper theme |
| ADHD | Distraction, attention drop-off | Focus Mode, Smart Chunking, Focus Structure simplification, Pomodoro timer |
| Low Vision | Contrast, font size, spacing | High Contrast themes, maximum spacing controls, TTS, large typography |
| General Reader | Comprehension, reading speed | Text Clarity simplification, Smart Chunking, Glossary, TTS |
Elu rewrites page content using a locally running quantized LLM. Three optimization modes address distinct cognitive profiles:
| Mode | Target Profile | Behaviour |
|---|---|---|
Text Clarity (textClarity) |
General / low reading fluency | Shortens sentences, removes filler, preserves paragraph structure |
Focus Structure (focusStructure) |
ADHD / attention challenges | Bolds key phrases, limits paragraphs to 1–3 sentences, adds scannable structure |
Word Pattern (wordPattern) |
Dyslexia / language learners | Enforces Subject-Verb-Object order, replaces idioms, avoids passive voice |
Each mode has five intensity levels (1 = light polish → 5 = 1st-grade rewrite), giving 15 distinct rewrite configurations. Levels are configurable globally (3-level compact mode) or with full granularity (5-level mode) via config.js.
Long articles are segmented into 150-word chunks (configurable range: 50–300 words). Each chunk is presented one at a time as a slide, with:
- Visual progress bar and "Chunk X of Y" counter
- Estimated read time per chunk
- Per-chunk bookmark toggle
- "Mark complete" action to track progress
- Session progress persisted in
chrome.storageby URL — reading position is restored on revisit
Focus Mode creates a distraction-free overlay on the current page:
- Configurable dim level for non-article elements
- Automatic hiding of ads, sidebars, comment sections, and related-content panels
- Animation and autoplay video blocking
- Optional ambient sound (brown noise, rain, cafe)
- Built-in Pomodoro-style countdown timer
- Toggle via popup or
Alt+Fkeyboard shortcut
A DOM tree-walker processes all visible text nodes and wraps the first half of each word in a <b> element with class bionic-highlight. This leverages the brain's ability to complete words from partial visual input, reducing cognitive load during scanning. The feature is fully reversible with removeBionicReading().
The TTS engine uses the Web Speech Synthesis API for cross-platform compatibility:
- Paragraph-level chunking for natural-sounding playback
- Variable playback speed (
tts-set-speed) - Voice selection from all system-installed voices (
tts-set-voice) - Play, Pause, Resume, Stop controls exposed in the popup
- Accessible via
Alt+Rkeyboard shortcut
A dblclick listener intercepts word selections across the page. Words present in a common-English dictionary are bypassed; uncommon or domain-specific terms trigger an API call to a dictionary service, with the definition displayed in a Shadow DOM tooltip that is visually non-intrusive and pointer-event isolated.
| Feature | Detail |
|---|---|
| 13 Colour Themes | Default, Dark Mode, Sepia, Cream Paper, High Contrast (B/W, Y/K, K/Y), Low Blue Light, Soft Pastel Blue/Green, Grey Scale, Blue Light Filter |
| OpenDyslexic Font | Bundled typeface designed to reduce letter-confusion errors common in dyslexia |
| Line Spacing | Adjustable 1.0× – 3.0× |
| Letter Spacing | Adjustable 0 – 10px |
| Word Spacing | Adjustable 0 – 20px |
On first install, a guided 3-step onboarding flow lets users self-identify their profile:
| Profile | Pre-configured Settings |
|---|---|
| Default | Standard spacing, system font, default theme |
| Dyslexia | OpenDyslexic font, Cream Paper theme, elevated line/letter/word spacing |
| ADHD | Dark Mode theme, level-5 simplification, moderate spacing |
| Low Vision | High Contrast theme, maximum spacing, elevated line height |
Profiles are immediately editable and all settings remain individually adjustable after onboarding.
Elu is designed with inclusion-by-design values:
- Screen reader compatibility: Notifications and status updates use ARIA live regions (
announce()helper) so screen readers surface state changes without visual focus interruption. - Keyboard-first navigation: All primary functions are operable without a mouse. Global shortcuts (
Alt+S,Alt+F,Alt+R) work from anywhere in the browser. - Contrast compliance: High Contrast and High Contrast Alt themes meet WCAG 2.1 AA contrast ratios. Yellow-on-Black and Black-on-Yellow variants meet AAA for users with achromatopsia.
- Font legibility: OpenDyslexic is bundled locally — no CDN dependency — ensuring it is available in offline environments.
- No motion by default: Focus Mode can suppress CSS animations and autoplay media, respecting users with vestibular disorders.
- Persistent preferences: All accessibility settings sync across Chrome profiles via
chrome.storage.sync, removing the need to reconfigure on each session. - Progressive enhancement: All non-AI features (bionic reading, focus mode, TTS, themes, spacing) function without model initialisation. AI simplification degrades gracefully with a clear "WebGPU not supported" message.
| Component | Technology |
|---|---|
| Extension platform | Chrome Extension Manifest V3 |
| Build tool | Vite 7 |
| On-device AI | WebLLM (@mlc-ai/web-llm 0.2.x) via WebGPU |
| Default model | Qwen2.5-0.5B-Instruct-q4f16_1-MLC (quantized, ~400 MB) |
| Content parsing | @mozilla/readability 0.6 |
| Markdown rendering | marked 17 |
| Speech | Web Speech Synthesis API |
| Storage | chrome.storage.sync |
| Fonts | OpenDyslexic (self-hosted) |
The extension ships no backend. Inference is delegated to a Web Worker running inside a Chrome offscreen document, using WebGPU for hardware-accelerated matrix operations. The offscreen architecture prevents WebGPU workloads from blocking the extension's UI thread.
Elu is designed to run on AMD Radeon hardware using the RDNA architecture's WebGPU compute capabilities. Every text simplification is a WebGPU workload dispatched directly to AMD GPU compute units via WGSL compute shaders.
| Integration | Detail |
|---|---|
| WebGPU Compute on RDNA | @mlc-ai/web-llm dispatches WGSL compute shaders to AMD GPU compute units for all LLM inference |
| MLC-AI Model Compilation | WebLLM uses the MLC framework to auto-tune and compile LLM operators (GEMM, attention) into optimised shaders for RDNA specifically |
| On-Device VRAM Management | Elu explicitly unloads and reloads the model via engine.unload() to be a responsible tenant on AMD integrated and discrete GPUs — critical on lower-VRAM Ryzen iGPU configs |
| AMD Ryzen AI — Forward Path | The same @mlc-ai/web-llm stack is positioned to route inference via WebNN to the AMD XDNA NPU, a natural hardware upgrade path for Elu |
Elu runs entirely offline. For users in low-connectivity environments, the AMD GPU in their existing device is the only compute available. By targeting WebGPU, Elu democratises on-device AI accessibility tools to anyone with AMD Radeon hardware — no cloud subscription, no data leakage, no ongoing cost.
| Requirement | Detail |
|---|---|
| Browser | Chrome 116 or later (stable, dev, or canary) |
| GPU | Any modern discrete or integrated GPU with WebGPU support |
| Disk | ~500 MB free space for the model cache |
| OS | Windows, macOS, or Linux |
Note: WebGPU support can be verified at
chrome://gpu. Elu detects WebGPU availability at runtime and disables AI features gracefully on unsupported hardware while keeping all other accessibility tools fully functional.
1. Clone and build
git clone https://github.com/Sri-Krishna-V/Elu.git
cd Elu
npm install
npm run buildFor development with hot-reload:
npm run dev2. Load the extension
- Open
chrome://extensions/ - Enable Developer mode (toggle, top-right corner)
- Click Load unpacked
- Select the
dist/folder produced by the build
3. Complete onboarding
The extension opens a guided onboarding page on first install. Select your accessibility profile and preferred AI model, then click Finish. The model will cache to disk on first use (one-time download).
| Shortcut | Action |
|---|---|
Alt + S |
Simplify the current page |
Alt + F |
Toggle Focus Mode |
Alt + R |
Toggle Read Aloud (TTS) |
Shortcuts can be remapped in Chrome at chrome://extensions/shortcuts.
- All text processing and AI inference run entirely on-device using WebGPU.
- No page content, reading history, or usage analytics are transmitted to any server.
- No account or sign-in is required.
- The extension only requests the minimum permissions required:
activeTab,storage,scripting,tts, andoffscreen. - The source code is fully open and auditable.
src/
background/ # Service worker: routing, lifecycle, keyboard commands
index.js
prompts.js # 15 system prompts (3 modes × 5 levels)
content/ # Injected into every page
index.js # Central message handler and orchestrator
smart-chunking.js
focus-mode.js
bionic.js
tts.js
glossary.js
offscreen/ # WebLLM engine host (WebGPU, Web Worker)
index.js
webllm-worker.js
options/ # Onboarding and settings page
popup/ # Extension action popup
common/ # Shared utilities
content-extractor.js # Readability + CSS selector extraction
config.js
logger.js
models/
chunk.js
focus-config.js
Contributions are welcome.
- Fork the repository
- Create a feature branch:
git checkout -b feature/your-feature - Commit your changes:
git commit -m 'Add your feature' - Push to the branch:
git push origin feature/your-feature - Open a Pull Request
Please ensure any new features maintain the extension's offline-first and privacy-first constraints.
Distributed under the MIT License. See LICENSE for details.
Accessible technology is not a feature. It is a right.