Shattering the Hallucination Barrier: Engineering a Self-Correcting Multi-Agent Compliance Pipeline

说明与用途

MAD 平台利用多代理架构与 RAG 对网页可访问性进行验证，提供确认的修复方案，降低小型企业的法律风险。

问题

Static code scanners for web accessibility produce noisy, unverified reports that leave small businesses vulnerable to ADA lawsuits.

结果

The MAD Platform delivers verified, production‑ready accessibility fixes for small businesses, reducing legal exposure and developer effort while operating at near‑zero cost through a serverless, scale‑to‑zero architecture.

工作流

• 1. Playwright headless crawl and screenshot capture.
• 2. Analyst Agent (gemini-3.5-flash-lite) runs deterministic, semantic, and multimodal visual checks per page.
• 3. Analyst flags potential violations.
• 4. Editor Agent (gemini-3.7-flash) receives flagged findings and performs RAG against a curated WCAG vector store.
• 5. Editor confirms or dismisses findings with documented reasons.
• 6. Reporter Agent compiles confirmed findings into Jira CSV/HTML reports.
• 7. Weekly Cloud Run background job analyzes dismissal logs, updates grounding rules for future scans.

证据

Standard rule‑based checkers function merely as noisy scanners…

来源

Instead of generating a static report, it deploys a deterministic, self‑correcting multi‑agent orchestration pattern…

来源