Research Agent for the FNR
Over 5,000 funded research projects at a German federal agency, answerable in plain language.
The Fachagentur Nachwachsende Rohstoffe is the German federal agency for renewable resources. It funds and supervises sustainability research across a wide range of fields, and everything it has funded is documented: more than 5,000 projects, each with an institution, a runtime, a task description and a final report, plus the reports and web pages around them.
All of it is public. Until now the only way into it was a keyword search.
That becomes a limit when the corpus is this size. A keyword search returns documents containing your word. It can’t tell you what twelve projects on the same question collectively found, or how the answer moved between 2015 and today. The FNR has more than 150 employees, and many of the people who need that kind of answer don’t have the scientific background to reconstruct it from the raw reports themselves.
What I built
An agent that searches the corpus iteratively. It reformulates its own queries, goes back in several times, and builds an answer across many projects at once with the sources attached.
The second half of it is register. The same question has a scientific answer, a strategic answer and a media answer, and they differ in length, depth and vocabulary. A press officer preparing a statement and a research officer preparing a funding decision start from the same 5,000 projects and need very different things out of them. The agent works out which of the three it is being asked for.
Alongside it I built an AI-assisted search across the FNR’s public website, which came out of a requirement the departments raised themselves.
And the data protection work, which was a substantial part of the project. A German federal agency can’t adopt a tool whose legal position is unclear, so I worked through the GDPR requirements and built them into the architecture from the beginning.
You can try the agent here.
Where it went
The agent landed well in the departments that used it, and it started a larger conversation.
An organisation of this size doesn’t adopt a tool in isolation. So together with the FNR I am now working out an AI strategy covering four areas: a central AI platform, internal systems, process automation, and public-facing applications. Each of them built from a concrete need somebody in the departments named, matched against what comparable public research organisations have already done about it.
I run the project independently, in direct coordination with the FNR.