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

Описание и применение

MAD Platform использует мультиагентную архитектуру с LLM и RAG для проверки веб‑доступности, выдавая проверенные исправления и снижая юридические риски для малых бизнесов.

Задача

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

Результат

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.

Рабочий процесс

• 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.

Доказательства

Standard rule‑based checkers function merely as noisy scanners…

Источник

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

Источник