An AI copilot that is not allowed to guess
Adding AI features to a dashboard builder, where the product has to catch a wrong answer before anyone sees it.
- React
- TypeScript

Outcome
- AI suggestions are checked before they reach anyone
- Four AI features built on the same rule
- Analysts can ask in Persian or English
The problem
Farakav’s dashboard builder lets analysts build their own reports. I am rewriting it in phases, and the brief also asked for AI: describe what you want, in Persian or English, and get it built.
A language model will happily give a confident answer that is wrong. In a reporting tool that is worse than an error, because people trust the chart that comes out of it.
What I did
I started from one rule: the AI suggests and the analyst decides. Nothing it produces takes effect until it has been checked and a person has accepted it. If a suggestion doesn’t hold up, the analyst gets a question instead of a wrong chart.
Four features now follow that rule. They help analysts build dashboards, write calculations and ask questions about what a dashboard shows.
The back end was new to me. I came to it as a TypeScript developer and worked on it with an AI coding agent, reviewing each phase and writing down the reasoning behind each decision.
What I would do differently
An early part of the AI work was built from a summary of the requirements instead of the specification itself, and it had to be redone. I now read the source document before building, however clear the summary sounds.