LLM Visualization Agent
Bounded agentic workflow that turns time-series forecast data into chart recommendations and reusable dashboards.
Brief
Built an LLM-based agentic application for visualizing time-series sales forecasts. The system retrieves forecast data, applies deterministic business logic, and gives the model only the schema and sample context needed to recommend an appropriate chart.
Problem
Forecast outputs were landing in analytical storage, but turning a new dataset into a useful dashboard still required engineers to inspect schemas, understand time-series shape, choose a chart, and wire the data into the interface.
Approach
I designed the workflow around a strict boundary: the LLM recommends the visualization, while the application owns data access, business logic, and rendering.
Highlights
- Built a function-call layer for retrieving forecast data and extracting schema plus representative sample rows.
- Scoped the reasoning layer so the LLM interpreted data shape and trends without generating application code.
- Mapped chart recommendations to reusable React, TypeScript, and ECharts components.
- Kept orchestration deterministic so the product flow stayed testable and explainable.
Outcome
Delivered a working prototype that established a repeatable path from forecast data to rendered dashboard. The prototype was selected for productionization after an executive demonstration.
Stack
LLMs, Google BigQuery, React, TypeScript, ECharts, deterministic orchestration, function calling.