Imported from jerrison/ada-compliance-auditor (AGENTS.md). Install upstream with npx skills add jerrison/ada-compliance-auditor. Copyright stays with the author.

AGENTS.md — ADA Compliance Auditor

AI-powered ADA accessibility auditor: photo upload → Gemini Vision analysis → ADA code mapping → cost estimation → structured report.

Quick Orientation

• Architecture: See ARCHITECTURE.md for full system design, data flow, and component responsibilities.
• Tech stack: Python/FastAPI backend, vanilla JS/Tailwind frontend, Gemini 2.0 Flash via Vertex AI, RocketRide pipeline.
• Package manager: uv (not pip). Use uv sync to install, uv run to execute.

Build & Run

# Install
uv sync

# Run (from project root)
cd backend && uv run uvicorn main:app --reload --port 8000

# Serves at http://localhost:8000

Environment

cp .env.example .env
# Required: GOOGLE_CLOUD_PROJECT=<your-gcp-project-id>
# Optional: GOOGLE_CLOUD_LOCATION=us-central1 (default)
gcloud auth application-default login

Project Structure

backend/main.py             → FastAPI app, routes, middleware
backend/gemini_client.py    → Gemini Vision API (direct + pipeline modes)
backend/pipeline_client.py  → RocketRide SDK client
backend/violations.py       → ADA code enrichment + cost mapping
backend/data/ada_codes.json → 18 violation types → ADA section references
backend/data/cost_estimates.json → 18 violation types → remediation costs
frontend/index.html         → UI (Tailwind dark theme)
frontend/app.js             → Upload, API calls, result rendering
frontend/style.css          → Severity colors, drag-drop states
pipeline/ada_auditor.pipe.json → RocketRide visual pipeline config

Code Conventions

Python (backend/)

• Async handlers for FastAPI routes
• Use logging module, not print()
• JSON responses for all API endpoints
• Imports: stdlib → third-party → local

JavaScript (frontend/)

• Vanilla JS only — no frameworks, no bundler, no build step
• Tailwind CSS via CDN
• Direct DOM manipulation

Architecture Invariants

1. Dual execution path: Always try RocketRide pipeline first, fall back to direct Gemini API. Never remove the fallback.
2. Data-driven violations: The 18 violation types are defined in backend/data/ada_codes.json and backend/data/cost_estimates.json. Adding a new violation type requires updating both JSON files AND the Gemini prompt in gemini_client.py.
3. Static frontend: No SSR, no build step. Frontend served via FastAPI's StaticFiles mount at /static.
4. No test suite yet: Validate changes manually by running the server and uploading an image.

API

Security

• .env is gitignored — never commit credentials
• Image data is in-memory only (not persisted)
• CORS allows all origins (demo/hackathon setting)
• Vertex AI authenticated via Application Default Credentials

Commit Style

• Small, logical commits — each reviewable independently
• PR descriptions explain why, not just what
• Link to relevant issue or spec when applicable