The latest articles on DEV Community by Pradyumna Roy (@pradyyy). https://dev.to/pradyyy https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4101228%2F20eba19f-bea6-466f-83f6-8f1165b94860.png https://dev.to/pradyyy en Pradyumna Roy Sun, 30 Aug 2026 10:05:04 +0000 https://dev.to/pradyyy/building-an-allergen-aware-ai-barista-using-google-cloud-adk-gemini-and-cloud-firestore-4mgp https://dev.to/pradyyy/building-an-allergen-aware-ai-barista-using-google-cloud-adk-gemini-and-cloud-firestore-4mgp <p>As conversational AI transitions from generic chatbots to specialized corporate agents, grounding language models in precise operational data has become a necessity. Recently, I built a serverless AI Barista application designed to act as an interactive coffee shop assistant. The architecture leverages the <strong>Google Agent Development Kit (ADK)</strong> alongside the high-efficiency <strong>Gemini 1.5 Flash</strong> model to drive intelligent customer conversations, manage contextual dialogue history, and surface accurate product recommendations.</p> <h3> The Backend and Security </h3> <p>To ensure production-grade security and data precision, the system avoids embedding raw text catalogs directly inside system prompts—which inflates response latency and transaction token counts. Instead, the backend utilizes <strong>Cloud Firestore in Native Mode</strong> as a scalable NoSQL vector database, implementing a flat composite vector index mapped to 768-dimensional embeddings via Google’s <code>text-embedding-004</code> model. The running application executes under a custom Identity and Access Management (IAM) service account conforming strictly to the Principle of Least Privilege, isolating database access roles from broader compute engine capabilities.</p> <h3> Application Hosting &amp; Validation </h3> <p>The application layer is wrapped inside an interactive, responsive <strong>Streamlit</strong> user interface and containerized for deployment on serverless <strong>Google Cloud Run</strong> infrastructure. During validation testing, the deployed RAG (Retrieval-Augmented Generation) agent successfully navigated out-of-menu safety traps without hallucinating, dynamically filtered allergen-specific requests (such as lactose intolerance exclusions), and seamlessly fetched dynamic menu additions (like a newly pushed 'Matcha Green Tea Latte') in real-time directly from the live production database.</p> googlecloud ai python serverless