The Lemonade Stand Assistant is an AI-powered customer service application designed to demonstrate safe and compliant interactions through multiple guardrails. Built by Anneli Sara Banderby and Cansu Kavili-Örnek, the project showcases how to deploy a conversational AI system that protects against harmful outputs, prompt injection attacks, and non-compliant language while maintaining a family-friendly experience. The fictional lemonade stand scenario provides a practical context where customers can inquire about products, ingredients, and pricing through a chat interface.
The system operates on three core security principles: the language model is treated as untrusted and all outputs must be validated, user inputs are untrusted and require validation, and detector triggering events are monitored and visualized for transparency. The architecture employs four distinct detection mechanisms working in concert. The IBM HAP Detector based on Granite Guardian monitors for hate, abuse, and profanity. A DeBERTa v3-based prompt injection detector identifies attempts to manipulate the AI assistant. The Lingua language detector ensures all interactions remain in English only. Additionally, a regex-based detector blocks specific competitor fruit references without requiring machine learning models.
The default language model is Llama 3.2 3B Instruct, though users can bring their own model endpoints through a Model as a Service configuration. The guardrails orchestrator coordinates all detectors using FMS Orchestr8 to evaluate both inputs and outputs before presenting responses to users. The application itself is built as a FastAPI-based web service providing the conversational interface.
Deployment occurs on Red Hat OpenShift with OpenShift AI and TrustyAI enabled. The solution includes comprehensive resource specifications, with the main LLM requiring 1 vCPU request and 4 vCPU limit plus 8 GiB to 20 GiB of memory and one NVIDIA GPU. The HAP detector needs 1 to 2 vCPU and 4 to 8 GiB memory. The prompt injection detector is the most resource-intensive at 4 to 8 vCPU and 16 to 24 GiB memory. The Lingua detector requires 1 to 2 vCPU and 2 to 3 GiB memory. Total cluster requirements are 7 vCPU request and 16 vCPU limit with 30 to 51 GiB memory and one GPU. Users can optionally enable GPU acceleration for individual detectors if additional GPUs are available.
The deployment includes an R Shiny monitoring dashboard that visualizes guardrail detections in real-time. The dashboard automatically fetches metrics every second and displays total request counts, input blocked counts, output blocked counts, approved request counts, and a detection breakdown by guardrail type with progress bars. This provides operational visibility into how frequently each detector is triggered and how the system is protecting interactions.
According to GitGenius activity tracking, the repository has minimal issue and pull request activity with a median response latency of zero hours across two tracked items. The most active contributor tracked is sarabanderby with ten recorded events. Bug reports and documentation requests represent the primary issue categories. The project is classified across multiple domains including AI Assistant, Business Simulation, Decision Support, Machine Learning, Profit Optimization, Predictive Analytics, Pricing Strategy, Inventory Management, and Strategic Planning, reflecting its multifaceted approach to demonstrating AI safety in a business context.
