Silv.IA: multi-agent analysis from one natural-language question

Context

The people who most need a churn or upsell analysis are rarely the people who can write the query for it. The request travels to a data team, waits in a queue, and comes back after the decision it was meant to inform has already been taken.

Problem

One large prompt over a whole dataset gives one plausible answer and no way to tell which part of it is wrong. And an analysis that ends in a document still needs somebody to notice the document.

Approach

Silv.IA takes a single natural-language question and triggers several LLM-powered agents, each responsible for one part of the answer (churn, upsell, general data analysis) and then acts on the result, sending an e-mail or a Discord notification without anyone writing a line of code.

Non-obvious decisions

• Decompose the question into specialised agents instead of enlarging one prompt. A narrow agent with a narrow contract fails in a way you can name and fix.
• End on an action, not on a report. An analysis that arrives as a message in the channel where the decision is taken gets used; one that waits in a dashboard does not.
• Make the plain-language question the only input. Anything that requires the asker to know the schema puts the data team straight back into the loop.

Outcome

A person with a question gets a multi-agent analysis and a notification in the channel they already work in, with no query written and no ticket opened.