Ask Data
An AI-augmented BI tool. Upload any CSV, ask questions in plain English, get charts and analysis back. Designed and built solo, deployed end-to-end on Vercel.
The hard part
Keeping the model honest about a schema it has never seen, and refusing the question when the data genuinely cannot answer it.
- React 19
- Vite
- Tailwind CSS
- Recharts
- Claude API
- Vercel serverless functions
01 / What it does
Replaces the “ask the analyst” workflow.
Ask Data is a natural-language interface for any tabular dataset. Drop in a CSV (sales, ops, finance, anything) and ask the data questions in plain English. The tool interprets the question, runs the right query, and returns a chart plus a written analysis.
The product was built to demonstrate what AI-augmented BI looks like in 2026. Your team does not need to write SQL or learn a BI tool. They ask questions, they get answers. For mid-market companies that have real data but cannot afford a full BI team, this is the pattern.
The tool is deliberately scoped: a single dataset at a time, no joins, no live database connections. That constraint is what makes it shippable solo and what keeps the user experience legible.
02 / Features
The five pieces that make it work end to end.
Drag-and-drop CSV upload
Browser-side parsing with type inference. The first 10 rows preview before commit; the user verifies columns and types before the dataset enters the analysis pipeline.
Natural-language query interface
A single text box accepts questions in plain English. The query is sent to Claude API via a Vercel serverless function, so no API keys are exposed in the browser. Claude returns a structured query plan plus a visualization recommendation.
Auto-generated charts
Recharts renders the response. Bar, line, and pie chart selection is driven by the query plan. The user can override the chart type after the fact; the choice is sticky for similar future queries.
Written analysis alongside the chart
Every response includes a 2 to 3 sentence written summary of what the chart shows. This is the layer that turns a visualization into an actual insight your team can paste into Slack.
Dark mode UI
Pure black surfaces with high-contrast text. Built for working sessions, not demos.
03 / Architecture
How the natural-language layer actually runs.
Browser
React 19 + Vite
Vercel Edge
Serverless fn
Claude API
Query plan JSON
In-memory data
PapaParse
Recharts
Bar / line / pie
- Frontend
- React 19 + Vite for fast dev iteration, Tailwind CSS for the design system, Recharts for visualization rendering. The app is a single-page client with hash-based routing for shareable analysis URLs.
- AI layer
- Claude API via Vercel serverless functions. The natural-language interpretation runs server-side, so no API keys are exposed in the browser. The system prompt is tuned to convert business questions into a structured query plan plus a chart-type recommendation. The plan is JSON-validated before execution.
- Data layer
- CSV parsing in the browser via PapaParse. Once parsed, the dataset lives in memory; queries run against the in-memory table using lodash-style filter and aggregate operations driven by the Claude-generated plan.
- Deployment
- Vercel, full edge for the static frontend, serverless functions for the Claude calls. Free tier comfortably handles development; production usage scales with token budget.
04 / Why this matters
For mid-market AI implementations.
Ask Data is the demonstration version of a pattern I would build for a client: an AI-augmented data interface that lets a non-technical operator ask questions of their data without going through an analyst.
For a mid-market company with a Salesforce export, an inventory CSV, or a finance reconciliation file, this pattern collapses a day of analyst time into a 30-second answer. The build cost is a fraction of a Tableau license and the iteration cost is near zero once the architecture is in place.
If your operations team currently lives in Excel, that is the conversation worth having. The first question is which of your workflows is the right one to migrate.
Next
Want something like this for your business?
Send me one paragraph about your actual data and workflow. If this is not the right tool for it, I will say so.