Data Analyst Agent

A conversational data-analyst agent built on n8n, Supabase (Postgres), and Google Gemini. Ask business questions in plain English; the agent translates them into schema-aware SQL, runs them read-only against Postgres, and returns a human-readable answer. Includes a synthetic retail dataset, two reproducible analysis workbooks, and a KPI dashboard built from the same data.

Educational / workshop project. The dataset (customers.csv, transactions.csv) is synthetic (Faker-generated), not real customer data.

What's here

• workflow.json — the populated n8n workflow (chat trigger → AI agent → Postgres query tool), ready to import and adapt to your own Supabase project
• workflow-template.json — the same workflow stripped down to a blank starting point
• customers.csv, transactions.csv — synthetic retail dataset (1,000 customers / 3,000 transactions) used to drive the workflow and the analyses below
• AGENT_ANALYSIS.md — reproducible SQL + methodology behind the two analysis workbooks below (top products/customers, and sales by location/channel)
• Top_Products_Customers_Analysis.xlsx, Sales_by_Location_Channel_Analysis.xlsx — the generated analysis workbooks
• sales-kpi-dashboard.pdf — a KPI dashboard summarizing the same dataset (preview below)

How it works

1. When chat message received — takes a natural-language question from the user
2. AI Agent (Gemini) — interprets the question against the known schema and generates SQL
3. Run SQL Query — executes the generated SQL (read-only) against Postgres
4. Response — the agent turns the result set into a plain-language answer

See workflow.json's embedded system prompt for the exact schema definitions and SQL formatting rules (PostgreSQL requires double-quoted column names here, since the source tables use mixed-case column names).

Setup

1. Import workflow-template.json into n8n.
2. Point the Postgres credential at your own database, loaded with customers.csv / transactions.csv (or your own data using the same schema — see AGENT_ANALYSIS.md).
3. Add your own Google Gemini API credential.
4. Fill in the system prompt (see workflow.json for a complete example) with your schema.

Limitations

Read-only queries only (no INSERT/UPDATE/DELETE), single database connection, and conversation memory is ephemeral (session-scoped). See AGENT_ANALYSIS.md for the production-hardening checklist (query validation, rate limiting, row-level security, etc.) if adapting this beyond a workshop exercise.