Tool-calling plant agents
Agents that turn requests into SQL, refresh dashboards, and raise alerts or tickets, with a person approving each action.
- Where
- Mercedes-Benz
- When
- 2025 – now
- Context
- Phase 2 of the plant roadmap
- Type
- Agents
SQLwritten from plain-English questions
Ticketsraised by agents, approved by people
5 shopscovered by the rollout
How it works
- 1RequestA question in plain English
- 2PlanAgent picks a tool
- 3ActSQL, dashboard or ticket
- 4ApproveA person confirms
- 5DoneQuery, view or work order
The problem
Phase 1 gave the plant one source of truth and dashboards. People still had to turn each question into a query, a dashboard change or a ticket by hand.
What I built
- Built agents that choose between tools: text-to-SQL, dashboard generation and refresh, alerts and maintenance tickets.
- Kept a human-in-the-loop approval step before any action reaches plant systems.
- Rolled out in Agile sprints with production, quality, maintenance and IT, with user training across 5 shops.
Stack
LLM agentsTool callingText-to-SQLLangChainFastAPIDatabricksPower BI