Arc
Arc lets blind and low-vision students study any paper, chart or video by ear, with keys and voice.
About the project
Many study pages are laid out for the eye. For a blind or low-vision student, the structure of the page and its figures are the hard part.
Arc rebuilds a public study page as a simple accessible “twin” with real headings. One input box: paste a link or say what you want. Hold Space anywhere and speak; arrow keys move; Escape stops speech. A cursor ring glides between sections. A guided figure tour speaks code-computed directions (“Move far down and to the left”), plays a soft chime, and names each part. A live panel shows the original page in a remote browser, following the cursor (read-only, never clicks). A YouTube link becomes an outline of what is said and what is shown on screen.
How it uses each technology
- Bright Data: three jobs. Web search from a spoken or typed request. Page fetch through Web Unlocker, with a saved-copy fallback. The Browser API remote browser for the live page view.
-
Gemini: Gemini Live (
gemini-3.8-live) for speech to text, tour speech and spoken answers.gemini-3.5-flash-litevision builds tours for unprepared figures (about 2 s) and outlines videos. - SubQ Snap (System One): turns each request plus the page’s headings and figures into a command and target with a probability, in about 0.2 s. Arc acts at 60% or more. Otherwise Gemini answers or asks what was meant: Gemini wraps Snap.
- Kokoro: local, open source voice that reads the page and confirms moves.
Parsing, ids, counts and all spatial directions are plain code, not a model.
Built with a software factory
One person directed AI coding agents in one day. An orchestrating session wrote a shared type contract and a brief with acceptance checks for each task. About a dozen agents (Claude Opus for the hard pieces, Claude Sonnet for the rest) built in parallel, each owning named files. Every piece passed an offline check and a real-browser test before commit.
Limits
Not yet tested with blind users or with a screen reader running. The live view understands MDPI’s page structure only. Video outlines are YouTube only, about 10 minutes at most. Model-made tour boxes can be wrong. The fast voice path uses Chrome’s speech recognition.
Next
Test with blind and low-vision students, and support more page layouts.
Built with
AI tools
Prize Categories Entered
- Submitted

Best Use of Gemini API
Gemini Live transcribes speech and speaks the figure tours and answers. Gemini vision builds a tour for any figure and extracts data from charts, with the numbers then computed in code. Gemini also outlines YouTube videos into what is said and what is shown on screen.
- Submitted
Best Use of Bright Data
Three jobs. The SERP API and Web Unlocker run the web search from a spoken request. Web Unlocker fetches pages that block plain requests. The Browser API drives a remote browser that shows the original page live and follows the student's cursor.
- Submitted
Best Demo Video Using HyperFrames
The demo video was composed and rendered with the HyperFrames CLI: the framed screen recording, captions, technology labels and end card are HTML compositions rendered to MP4.
- Event prizeSubmitted
Best Overall Hack
Arc lets a blind or low-vision student study any paper, chart or video by ear. They speak, Arc finds the page, reads it, and guides them through each figure part by part with a chime on arrival. It combines four technologies, each doing the one job it is best at, and every piece was built and tested in one day. It is not yet tested with blind users, and we say so.
- Event prizeSubmitted
Best Use of System One Models
SubQ Snap decides every command. It takes what the student said plus the page's headings and figures, and returns the action and target with a probability, in about 0.2 seconds. Arc only acts when Snap is at least 60% sure. Otherwise Gemini answers or asks what was meant. A fast decision model lets Arc refuse to guess, which matters when the user cannot see a wrong move.
- Event prizeSubmitted
Cleanest Use of a Software Factory
One orchestrating session wrote a shared type contract and a brief with acceptance checks for each task. About a dozen coding agents built pieces in parallel, each owning named files. Every piece had to pass its own offline check and a real-browser test before it was committed, and releases were built from clean copies of a commit. The README documents the process.