About datavisual.studio

datavisual.studio turns a dataset into a living, editable dashboard. Upload a file or connect a SQL database or REST API, and you get charts and metrics you can refine by chatting in plain language. From the same place, a council of AI models can research a question against the live web and pin its findings — cited — onto the dashboard. One “Update” later re-pulls the data and re-runs that research, and tells you exactly what changed.

It is free. You bring your own AI keys, and your data stays on the platform's own server.

I'm Mohammed Isa (@4mohdisa), a software engineer. datavisual.studio is a solo project — product, backend, frontend, infrastructure and design. The source is open under the MIT license.

• Build a dashboard instantly from CSV, Excel, JSON, or a live SQL/REST connection.
• Edit it by chat (“add a bar chart of revenue by region”) or by hand.
• Ask questions of your data and get grounded, defensible numbers back.
• Run a multi-model research council that reads the live web and writes a cited report.
• Monitor it: one update surfaces value deltas and fresh sources since last time.
• Share a read-only public link, or export to PDF / HTML.

The stack is a Next.js frontend on Vercel talking through a single authenticated proxy to a FastAPI backend on AWS, with Clerk for identity. There is deliberately no database: all state is JSON on disk. For a single-tenant, single-replica product that removes an entire operational surface and makes the security model tractable — one directory is the trust boundary.

The decision I care about most: the language model never does arithmetic. It translates a question into a JSON query spec; a deterministic Python engine executes that spec and phrases the answer from the result. A measure that's a level (MRR, headcount) is never summed across time; a “how many” that the model accidentally split by a category it wasn't asked about is collapsed back to the total. Every number the product shows can be defended — because a person, or a test, can re-derive it. Charts carry a deterministic text alternative for screen readers for the same reason: the summary is generated from the plotted numbers, not narrated by an LLM.

A fuller write-up of the architecture and its invariants lives in the repository. If you want to see how it fits together, that's the place to start.

Open the studio →