GlassBox is the investigation dashboard for Inhibitor. It takes raw intervention logs, shows you exactly what your AI did wrong and what was corrected, finds patterns across incidents, answers questions through a built-in chatbot, and generates audit reports you can hand to a regulator.

Most audit tools give you charts and tables. Glass Box gives you understanding. It works in three layers. First, it parses Inhibitor logs and tags every intervention by type, severity, and policy, so you see the full picture without reading a single CSV row. Second, it has a built-in chatbot where you can ask questions in plain English like "Why was this blocked?" or "Show me similar incidents" and get context-aware answers. Third, it runs pattern detection across all interventions, clustering similar events, surfacing repeated failures or suspicious behavior, and generating insight summaries. The standout feature is a one-click Risk Report generator that produces a full summary of system behavior, key failures, and actionable recommendations, ready to share with stakeholders or attach to a compliance filing.

Next.js 14 with App Router for the frontend and backend, TypeScript across the full stack, Tailwind CSS and shadcn/ui for the interface, Recharts and D3.js for charts and the risk correlation graph, Framer Motion for animations. Backend runs on Next.js API Routes with 7 endpoints handling file upload, analysis, chat, report generation, dataset management, and request drill-downs. SQLite through Prisma ORM for the database with zero external setup. Anthropic Claude API on the server side for the investigation chatbot with streaming responses, automated executive summaries, and compliance report narrative generation. PapaParse for CSV log parsing, PDFKit for server-side PDF audit report generation. 4 bundled Inhibitor audit datasets covering insurance, healthcare, fintech, and edtech with around 86,000 pipeline events total.

https://github.com/rootwij/inhibitor-lab