Data & analyticsStatus: Pilot
Controlling cockpit with an AI data analyst
Challenge
Finance and sales figures were spread across several databases and hard-to-read fact tables. Management needed a consolidated live view and answers to ad-hoc questions without writing SQL.
Technology
- Python
- FastAPI
- SQL Server
- LLM-Agent (Text-to-SQL)
- MCP
- React
- Docker
Requirements
- Strictly read-only access to production databases
- P&L, budget/actual/prior year, cost centres and forecast
- Questions to the data in natural language
- Offline operation on a VM with a data snapshot
- No credentials or business data in the code
Approach
- 01Backend with a technical read-only guard on the database connection
- 02Aggregation layer for a sign-correct P&L from the data cube
- 03AI agent with a text-to-SQL tool and schema briefing
- 04Dashboard with 14 chapters, global filters, bookmarks and PDF/CSV export
- 05MCP server for AI clients; Docker bundle with a local database and cent-exact verification