RedDial

An adversarial simulation and evaluation harness for conversational AI agents. It spins up difficult customer personas (angry escalators, prompt injectors, policy probers), runs them against a target agent in parallel, and grades each transcript with deterministic decision-tree judges for task completion, tone and policy, and groundedness against the target's own documents. Built on LangGraph for parallel multi-agent orchestration and LangChain for the model layer, with a small RAG pipeline that catches hallucinated prices and policies, and a self-contained HTML report that draws each judge's decision path. TypeScript, tested with Vitest, on npm.