Sergio AlbaBogotá, Colombia

Lab 03Generative AI

Generative Dashboard

Turns a natural-language question into an interactive, verifiable dashboard.

Explore the dashboardHow it works

marketing_demo v1.0.0 · synthetic data

01The problem

Where is my marketing spend losing efficiency this quarter?

A question contains more structure than it seems: objective, period, dimension and metrics. Answering it means identifying that structure and calculating every number from the data, without asking AI to invent numbers.

02What it does

  1. 01

    From question to structure

    Identifies intent, available metrics and dimensions, and turns them into a structured analysis plan.

  2. 02

    A clear role for AI

    It can interpret intent, identify metrics and dimensions, propose visualizations and change the structure. It cannot invent data or columns, execute arbitrary code, alter sources or bypass validation.

  3. 03

    Verifiable numbers

    Every number shows its registered formula, its source values and the SQL query, and passes five checks before it is displayed.

  4. 04

    Seven possible designs

    One conversation, seven designs: spend and CPL trend, performance by device, monthly comparison, acquisition by channel, device detail, executive summary and revenue by campaign.

03How it works

  1. 01

    Question and interpretation

    A user question. Intent, available metrics and dimensions are identified.

  2. 02

    Plan

    A structured object defines the analysis. An invalid plan leaves the dashboard unchanged.

  3. 03

    Calculation

    The engine executes a limited query catalog. AI never executes code or arbitrary SQL.

  4. 04

    Validation

    Fields, period, formula, denominator and sufficient data are checked. A zero denominator returns N/A, never 0.

  5. 05

    Visualization

    The validated result feeds the dashboard components. History keeps every version so you can undo, redo or restore it.

04Data and limits

Natural language. Bounded operations.

Guided demo: it supports seven catalog commands, and each one selects a complete plan, using deterministic rules and without calling an AI model. The data is marketing_demo: 180 fixed synthetic observations from April to September 2025, in USD, which do not represent real campaigns.

Built with

  • React
  • SQL