Lumina Read

Dyslexia-aware text simplification for the web.

Lumina Read Lumina Read is a Chrome extension that rewrites hard-to-read text into plainer language, directly on the page, for people with dyslexia. WHAT IT DOES You select any text on any web page and let go of the mouse. A box appears over what you selected containing a simpler version of it, written out sentence by sentence as it is generated so you are not waiting in front of a blank box. Clicking anywhere dismisses it. The box stays anchored to the text it belongs to rather than following your cursor around, and it sizes itself to the length of what you selected. Two sliders in the popup control how far it goes. One sets the reading level. The other sets how hard it should work to avoid long words and letters that are easy to confuse with each other, like b and d, or p and q, which is a specific difficulty for dyslexic readers that general "simplify this" tools ignore. Alongside the rewriting there are three reading aids: dyslexia-friendly fonts (Atkinson Hyperlegible and OpenDyslexic), increased letter spacing and line height, and a line focus mode that dims the whole page except the line you clicked on, so you can hold your place. WHO IT IS FOR Primarily people with dyslexia, which affects a significant share of the population. But the specific problem it solves is narrower than "reading is hard". It targets the text nobody gets to opt out of: medical instructions, rental agreements, insurance policies, employment contracts and government benefit forms. That writing is dense by convention rather than by necessity, and it is exactly where misreading something carries a real consequence. Existing accessibility tools mostly change how text looks. Changing the font of a complicated sentence does not make the sentence any less complicated. Lumina Read changes the words. HOW IT WORKS The AI model runs entirely on your own device. It downloads once, about 700 MB, and then runs inside your browser on your graphics card using WebGPU. There is no server, no account and no API key, which means no page content ever leaves your computer. That was a deliberate architectural decision rather than a convenience: people use this tool on medical and financial pages, and a tool that shipped those pages to someone else's server would not deserve to be used on them. Getting a language model to live inside a Chrome extension turned out to be the hardest engineering problem. Chrome shuts down an extension's background worker after roughly 30 seconds of inactivity, which would throw away the loaded model over and over. Running it inside each page instead means every open tab loads its own copy and consumes over a gigabyte of graphics memory. The solution was an offscreen document, a hidden page the extension keeps alive, so one copy of the model is shared by every tab. ACCURACY The failure mode that matters for an accessibility tool is not a clumsy rewrite. It is a fluent, confident rewrite that has quietly dropped a dosage or a deadline, because the reader using this tool is the one least able to notice. Small models that can run on a laptop do this readily. In one measured case the model turned a 1,200 character passage into 197 characters and lost all eight numbers in it. So Lumina Read checks its own output before showing it to you. Every rewrite is compared against the original. If it invents a link, invents an email address, changes a number, answers as a chatbot instead of rewriting, or summarises rather than simplifies, the result is discarded and you keep the original text. If a single figure goes missing, you still get the rewrite, shown with a note naming exactly what to check against the source. The reasoning behind that split is what the reader can do about it: a fabricated fact is undetectable, while an omitted fact is still sitting on the page right behind the box. The same checker also defends against prompt injection. Page text is untrusted input, so a page can contain instructions aimed at the model, and the output gets rendered back to the reader in the extension's own styling. BUILT WITH TypeScript, React and Tailwind CSS for the interface, the Plasmo framework for the extension build, and WebLLM on WebGPU for on-device inference. 597 automated tests across 23 suites cover the logic, the tooltip geometry, the accuracy checker and the message protocol.