An illustration tracing a loop from a honeycomb crystal lattice through a stacked layered material, a device array, a circuit grid, and a neural network, with arrows running in both directions.

Agentic AI for a Physics Lab

The short version

A research lab's knowledge lives in scattered places: hundreds of paper PDFs, a wiki of procedures, and instruments driven by measurement software. I initiated and lead an effort in the Kläui Lab to make all three usable by AI agents without sacrificing trust: a literature assistant that cites the exact sentence behind every answer, a queryable lab wiki, and instrument control that agents can operate safely. One strand is running end to end; the other two are in early stages.

Role
Initiator and lead builder
Period
2026–present
PythonRAGMCPPostgreSQLLLM agents

This is the umbrella for a lab-wide digitalization effort I started and lead in the Kläui Lab at JGU Mainz. It grew out of two tools I had already built for my own research: CryoSoft, the measurement software that runs the lab's cryogenic experiments, and TracEx, a tool that extracts measured values from scientific papers with a traceable source for every number. This initiative turns their ideas into shared lab infrastructure.

The problem

The knowledge a physics lab runs on is fragmented. What has been measured and published sits in hundreds of paper PDFs. How the lab actually works, its procedures, recipes, and equipment know-how, sits in a SharePoint wiki. And the experiments themselves are driven by measurement software that only a human can operate. A generic AI chatbot helps with none of this reliably: it cannot see internal material, and when asked about literature it invents citations. Each strand of this project makes one of those knowledge stores usable by AI agents, with answers that can be checked rather than taken on faith, and with lab data staying on lab infrastructure.

The approach

Literature RAG is the furthest along. It parses the group's paper corpus (234 papers) with full provenance down to page and character position, indexes it for hybrid semantic-plus-keyword search, and serves it to AI agents through an MCP server whose answers trace back to the exact supporting sentence in the exact paper. The parsing-and-verification philosophy comes directly from TracEx: stamp provenance anchors on the text before any AI model is involved, then verify every model claim against its cited anchor afterwards.

Wiki RAG applies the same retrieval approach to the lab's internal SharePoint wiki, so that a new lab member can ask how a procedure works and get an answer grounded in the lab's own documentation. This strand is early: current work is on programmatic access to SharePoint through Microsoft's Azure AD and on a parsing pipeline for the wiki's pages.

Agentic instrument control extends CryoSoft, the layered cryostat measurement system I built, toward operation by AI agents: an agent harness for adding drivers and troubleshooting already exists on the CryoSoft side, and MCP tools exposing measurement status are in progress. This strand is at the design and literature-review stage, informed by the emerging self-driving-laboratory field.

Beyond the software, the initiative includes the organizational side: convening the lab's digitalization meetings and arranging GPU hosting with the university's data center so the tools can be served to the whole group.

Status

Literature RAG runs end to end on the lab's GPU workstation; the next step is wiring it into the university's internal AI chat service, currently pending a hosting request. Wiki RAG and agentic instrument control are early-stage, and this page will say so until that changes.