Gawri Analytics

Automated exploratory data analysis using Large Language Models (LLMs) often results in unreliable, non-deterministic outputs, making fully autonomous dashboard generation unfit for deployment-ready environments. This paper introduces Gawri Analytics, an agentic framework that addresses this reliability gap by strictly separating intent detection, visualization planning, and constrained execution. In contrast to fully generative pipelines, our techniques enforce structural validation before code execution and restrict runtime behavior through specific safety guardrails. We evaluated it across multiple datasets and a diverse set of natural language queries covering analytical, visualization, and adversarial categories; the framework achieves a 92% execution success rate while demonstrating strong resilience to adversarial inputs and substantially reduced latency compared to fully generative baselines. These results suggest that joining deterministic routing with constrained LLM-based planning offers a feasible path toward safe, low-latency automation.