Overview
In large companies, employees often have access to huge amounts of data – spread across various internal systems and platforms. The challenge: gaining relevant insights quickly, without having to work through dashboards and databases. This is exactly where this project comes in.
For a large German automotive group, we developed an internal AI-powered analytics tool based on a multi-agent architecture. Through an intuitive chat interface, employees can ask questions in natural language. In the background, several specialized AI agents orchestrate queries against various data sources – including Databricks – and deliver processed answers, concrete recommendations for action, and proactive risk warnings.
The result is a tool that significantly accelerates data-driven decisions and gives even employees without deep technical knowledge access to complex analyses.
My Role
As a fullstack developer, I was part of a larger development team from September 2025 to March 2026 and took over the lead of the frontend development. This meant both technical responsibility for architecture decisions in the frontend and close coordination with the backend team around the agent logic and API interfaces. Beyond that, I actively contributed to the backend development, in particular the agent orchestration and the integration of data sources.
Tasks
- Frontend development lead – Technical leadership and architecture decisions for the Next.js-based frontend within a larger team
- Chat interface with real-time streaming – Design and implementation of the central user interface through which agent responses are streamed in real time
- Data visualization – Presentation of complex analysis results and risk indicators in understandable, interactive formats
- Backend development with LangGraph – Development of API endpoints and implementation of the agent orchestration in the Python backend with FastAPI
- Data source integration – Connection of various internal systems, including Databricks, as the knowledge base for the AI agents
Technical Highlights
Streaming architecture for agent responses: Since multi-agent processing can take several seconds depending on the complexity of the request, seamless real-time streaming of the responses was essential. Partial results from the individual agents are transmitted directly to the frontend via Server-Sent Events, so users can follow the progress live.
Multi-agent orchestration with LangGraph: The various agents – for example, for data queries, risk analysis, and recommendation generation – are orchestrated via LangGraph as a directed graph. This makes it possible to model complex workflows in which agents build on each other or work in parallel.
Secure integration into the corporate infrastructure: The entire system runs on Azure and uses Azure OpenAI as the LLM backend. The connection to Databricks and other internal data sources is made via secured interfaces that meet the group’s security requirements.