Shyamnath Sankar

Project 03 of 5Agentic database analysis2026

SelectStar

Ask the data. See the answer.

Role
Creator
Year
2026
Stack
Next.js 16, TypeScript, Prisma, Vega-Lite
Links
SelectStar landing page: Talk to your database in plain English, with a connection-string form and a diagram of router, SQL, EDA, viz and ML agents
Paste a connection string and ask.

The brief

Analysis should start with a question, not a query.

SelectStar lets a person connect a database and ask what they need in everyday language. The system decides whether the answer calls for SQL, exploration, a chart or a small machine-learning model, and shows every statement it runs.

The system

RouteAnalyzeExplain
  1. 01

    Route

    Classify the intent and pick the specialist agents.

  2. 02

    Analyze

    Query, profile, chart and forecast, in parallel where it can.

  3. 03

    Explain

    Stream a clear answer with the SQL and artifacts beside it.

The build

A graph of specialists, not an open-ended loop.

A custom TypeScript graph orchestrator runs inside Next.js, with no Python service. Specialist agents handle SQL generation, EDA profiling, Vega-Lite charting and pure-TypeScript ML, then a synthesis agent streams the result. Writes are intercepted and need explicit confirmation.

  • SQLite and PostgreSQL, MySQL-ready architecture
  • Read-only by default, writes gated per statement
  • SQL, tables, charts and models as canvas artifacts
  • Every statement shown before and after it runs
SelectStar chat asking how many orders there are by status, with the generated SQL and the result table on the canvas
Every answer shows the SQL that produced it.
SelectStar answering a request to chart the top eight products by total sales, with the generated SQL and its results on the canvas
Ask for a chart and the agents write and show the query first.
SelectStar clustering products into three groups by price and stock, with the model summary in the chat
Small ML models run in pure TypeScript.

The outcome

Safe: by default

A deterministic graph replaces the ambiguity of an agent loop: read analysis branches into parallel EDA, charts and ML, and nothing touches the data without a confirmation.