Speak to data: ask the dashboard

Context

I have built dashboards from single pages to large semantic models. The larger the organization, the more complex the questions it needs answered. A dashboard shows the agreed numbers, and follow-up questions usually go to an analyst. This project gives the dashboard an analyst of its own: the user asks a question and gets the answer in real time. I am building it on a KPI dashboard for the SaaS industry.

What I did

• First the semantic model and the metrics in dbt, with tests.
• Then a chat on the front end: a plain-language question becomes a call to a deterministic dbt metric, and the answer comes back in plain language.
• At period close, the model writes a text summary of the report and answers questions on any detail. Useful for companies with a branching hierarchy, to see where attention is needed.
• Metrics are versioned YAML contracts and staged by company stage, not one Rule of 40 number.

Result

• The model handles language, never arithmetic. I can demo it live.

Stack

dbt, DuckDB, Postgres, Python, Next.js, Claude API