Aria
AI Website Accessibility Auditor
Demo Video
Screenshots
Overview
An AI-powered accessibility auditor that combines deterministic axe-core rule-based analysis with Claude Vision's visual reasoning to surface issues automated scanners structurally cannot detect.
Technical Details
Combined axe-core (injected into the live DOM via headless Chromium) with Claude Vision analysing the rendered screenshot — surfacing issues automated scanners structurally cannot detect: weak visual hierarchy, cramped touch targets, text-over-image contrast failures, and misleading reading order.
Tagged every finding by source — deterministic axe-core (certain) vs. AI-inferred (probable) — and weighted them separately in the severity score, so model output is never presented with the same authority as rule-based findings.
Ran headless Chromium serverless via puppeteer-core + @sparticuz/chromium; constrained the Claude vision prompt to emit only visible issues as structured JSON with WCAG 2.2 criterion references, suppressing hallucinated findings.