How kartik-idk/accessaudit works
AccessAudit is a research prototype for AI-assisted accessibility remediation, with an interactive audit workflow and offline model benchmarks. The documented end-to-end path starts with browser/Axe violations, localizes them to source, obtains a constrained proposal, validates it, patches source, and verifies the result. The README documents this sequence; sampled code confirms the proposal gatekeeper and source write, but does not expose all runtime wiring, so unsampled stages use the enclosing scripts path and uncertain calls are omitted.
Main entry points: scripts Browser Runner · main.tsx Mode Router · PrototypeUI.tsx Prototype UI · pipeline_v6.js Proposal Gatekeeper
Components 18
User Experience
- Mode Router
src/main.tsx - Prototype UI
src/components/PrototypeUI.tsx
Audit and Localization
- Audit Service
scripts - Browser Runner
scripts - Axe Scanner
scripts - DOM Violation
scripts - Source Provenance
scripts - Localized Context
scripts
Proposal and Patching
- LLM Proposal
scripts - Proposal Gatekeeper
scripts/pipeline_v6.js - Source Patcher
scripts/pipeline_v6.js - Target Source
Verification and Research
- Build Check
scripts - Axe Verification
scripts - Model Benchmarks
experiments/task30_base_vs_finetuned/run_task30.py
Other components
- Engineer
- Ollama
- Playwright
Connections 18
- Engineer → Mode Router: select mode
- Mode Router → Prototype UI: renders
- Prototype UI → Audit Service: requests audit
- Audit Service → Browser Runner: runs audit
- Browser Runner → Playwright: drives browser
- Browser Runner → Axe Scanner: scans page
- Axe Scanner → DOM Violation: reports violation
- DOM Violation → Source Provenance: localizes source
- Source Provenance → Localized Context: extracts context
- Localized Context → LLM Proposal: provides context
- LLM Proposal → Ollama: requests proposal
- LLM Proposal → Proposal Gatekeeper: submits proposal
- Proposal Gatekeeper → Source Patcher: accepts proposal
- Source Patcher → Target Source: writes patch
- Target Source → Build Check: builds source
- Build Check → Axe Verification: enables verification
- Axe Verification → Axe Scanner: checks outcome
- Model Benchmarks → Ollama: evaluates models