Trendyol's Data Warehouse team built a conversational AI agent using the Agno framework that lets anyone query GCP billing data in plain language. The agent pre-loads metadata (projects, services, SKUs) to avoid LLM hallucinations, uses fuzzy matching to resolve informal user terms to exact GCP names, and generates production-grade BigQuery SQL following strict guardrails: no LIKE wildcards, unique CTE aliases, mandatory cost/credits/net-cost triple, and dynamic date functions. A custom BigQuery tool executes queries safely. Data governance is embedded directly in the agent's instructions, enforcing project-level access control with clear denial messages. Key lessons include the value of pre-loaded context, SQL guardrails as reliability features, and the necessity of fuzzy matching since users rarely type exact service names.
1 Comment
Sort: