Lemonade Stand Assistant is an AI-powered customer service assistant that enforces safety guardrails for compliant, family-friendly interactions.
The tool addresses the challenge of deploying language models safely in customer-facing applications. It treats both the LLM output and user input as untrusted, validating each through multiple detector models before allowing any interaction. The system uses Llama 3.2 as its default language model, protected by three specialized detectors: the IBM HAP Detector monitors for hate, abuse, and profanity; a prompt injection detector identifies manipulation attempts; and a language detector ensures English-only communication. A regex-based detector adds an additional layer by blocking specific competitor brand names. These guardrails are orchestrated to evaluate both inputs and outputs, with all detector triggers monitored and visualized through a dashboard.
The tool suits teams building customer service chatbots or conversational interfaces where safety and brand compliance are critical. It works well for businesses that need to demonstrate responsible AI deployment with visible monitoring of safety violations. The interactive demo allows you to test how the system responds when users attempt to abuse it or inject prompts, making it useful for understanding guardrail behavior before deployment.
Work in the issue tracker centers on bug fixes and documentation improvements.
