The service
AI Agents & Multi-Agent Systems
For tasks that need several data sources, calculation steps and tools. With clear boundaries and human approval.
Multi-agent system for a German automotive group.
An internal AI tool that answers analysis questions across several data sources. Specialized agents fetch the data, an orchestrator assembles the answer – live in the chat, with charts.
- Context
- German automotive group
- Service
- AI Agents & Multi-Agent Systems
- Role
- AI Engineer & Fullstack
- Period
- Since 09/2025, ongoing
- Setup
- In the client’s existing team
01 — Starting point
The data was there. The path to the answer wasn’t.
The question behind it
“Why do we have to hunt for data before we can answer any analysis question?”
The relevant data was spread across several sources. For a single analysis question, specialists had to know where everything lived, build queries, combine the results and prepare them. That took time – and depended on the few people who knew their way around every system.
02 — Solution
Ask questions instead of searching for data.
A chat up front that anyone can use. Behind it, a system of specialized agents.
- 01
One chat instead of many queries
Specialists ask their question in natural language. The system decides which data it needs – not the person who would otherwise have to gather it.
- 02
An orchestrator that plans
It breaks the question down into subtasks and hands them to the right agents.
- 03
Specialized agents per data source
Each agent knows its source and has only the tools it needs for it. Connected through a dedicated backend, including to the group’s data platform.
- 04
Answers you can see
Results are streamed live and shown as charts right in the chat.
03 — Architecture
From the chat to the data source.
Simplified and anonymized. Internal systems and data sources are deliberately not named.
A · Interface
Chat
Analysis question in natural language. Answer with live streaming and charts.
Next.js · TypeScript
B · Orchestration
Backend
Receives the question and streams the answer back.
Python · FastAPI
Orchestrator
Breaks down the question and distributes subtasks.
LangGraph · Azure OpenAI
C · Agents
Agent · Source 1
ToolsAgent · Source 2
ToolsAgent · Source n
ToolsOne specialist per data source, with only the tools it needs.
D · Data
Data platform
The group’s existing data platform.
Databricks
Other sources
Internal systems, anonymized.
← E · Answer
Partial results are combined and streamed live into the chat, with charts.
04 — My role
The whole system, not just the prompt.
AI Engineer and fullstack developer in the client’s existing team, since 09/2025.
- 01
Agents
Orchestration of the specialized agents: their tasks, their tools and how they work together.
- 02
Backend & data
API for the agents’ communication and the connection to the data sources.
- 03
Interface
Chat with real-time streaming and data visualization.
- 04
In the team
Following the group’s processes, platforms and approvals.
05 — Result
What has changed.
Concrete numbers are under NDA. I’m happy to tell you more in a conversation, as far as I’m allowed to.
- One chat instead of queries
- Analysis questions are asked directly. Nobody needs to know beforehand which system holds the data.
- Several sources, one answer
- The agents combine data from different systems and deliver one coherent answer with a chart.
- At the group since 09/2025
- Ongoing development in the existing team, on the client’s platform and to the client’s standards.
06 — Tech
The stack.
For technical readers, teams and agencies.
- AI & agents
- LangChain, LangGraph, Azure OpenAI, Azure AI Foundry
- Backend
- Python, FastAPI
- Data
- Databricks
- Interface
- Next.js, TypeScript, real-time streaming
Technical details
What I built in detail, as far as it can be made public.
Agents as a directed graph
The agents – for example for data queries, risk analysis and recommendations – are orchestrated with LangGraph as a directed graph. That way, agents can build on each other or work in parallel.
Partial results via streaming
Depending on the question, processing takes several seconds. Partial results from the individual agents go straight to the frontend via server-sent events, so users see the progress live.
Frontend lead
Technical responsibility and architecture decisions for the Next.js frontend, in close coordination with the backend on agent logic and API interfaces. Plus the presentation of analysis results and risk indicators as interactive visualizations.
In the group’s infrastructure
The system runs on Azure, with Azure OpenAI as the model backend. Databricks and other internal sources are connected through secured interfaces that meet the group’s security requirements.
Internal systems, data sources and metrics are under NDA.
07 — Next
More like this.
How I build AI agents
Similar project
AI assistant for the knowledge of a school
An internal wiki and a legal reference work of around 1,400 pages as a chat with sources. Built alone, through to operation.
Start a project
Your data, your questions, one chat?
Tell me in two or three sentences what it’s about. You’ll get an honest assessment within one business day – even if it says you don’t need AI at all.