Lorenz Ehrlich

Could you explain what you worked on during your Final Project and how that work has evolved into what you do now at Telespazio/ESA? What real-world impact do you expect your project to have?

My Final Project focused on developing a Large Language Model powered Interface designed to emulate the XMM Newton Helpdesk, a man powered Q&A service offered by ESA where astronomers can send in questions, errors, and general inquiries concerning their work with XMM Newton related data. The idea was to take 20 years worth of Tickets in which experts had answered user questions and fine-tune a language model to act as a XMM Chabot. Other aspects of the project included building a customized Retrieval Augmented Generation (RAG) system, designed to grab relevant information concerning a question from XMM related documentation. The process of building focused heavily on the exploration of various methods, including the fine-tuning of embedding models, hybrid retrieval systems, rerankers, text chunking strategies, metadata filtering, query-document routing and many more. 

My work at ESA now has continued to focus on the XMM Chatbot, evolving from a static question-answer system to a fully-fledged autonomous agent which can directly perform astronomical data processing tasks for users. The technical details of this exactly are slightly too complex to cover in a document as this. In terms of real world impact I think there are 2 main ideas. Firstly, the system does intend to serve real astronomers by aiding them in their research process, and is currently live to a limited user group within the ESA office in Madrid. Secondly the project aims to serve as a foundation/exploration of LLM agents for Astronomy, attempting to explore what works, what doesn’t, what best practices may be, and how we can create these systems at scale for missions outside of XMM. 

What was your first visit to the ESA site in Madrid (ESAC) like while you were working on your Final Project for the Bachelor in Computer Science and Artificial Intelligence?

It was lovely. Of course I was sort of enamored with the idea of going to an ESA office, the name alone carries a lot of weight. It felt exciting to go visit the frontier of Astronomy. The office generally also lies in a scenic and beautiful location. I was also excited to meet my supervisor, Aitor Ibarra, in person, and meet the team behind XMM Newton. They were all very welcoming and interested in what I was working on, the whole atmosphere in the office is rather inviting. 

How would you describe your collaboration with the team—scientists, engineers, and data specialists?

In one word, Vital. While I have the ML/AI knowledge to undertake a project like this, I had not had any experience with Astronomy. Specifically Astronomical Data Analysis Software. So developing a project focused on that would be almost impossible without the native knowledge of the XMM team, specifically my supervisors who have been nothing short of spectacular in their help within the project. They bring in ideas on features, they can tell me exactly what APIs are available for implementation, they can give me advice on how to structure something so that the end user gets what they want. As its intended to be a user facing project having expert users by my side to advise me was extremely important and valuable. Additionally collaboration between departments is also valuable, the data science team has been a big help in setting up the hardware powering the system. And generally there are far more people that have contributed to the project in some way, big or small, than they realise. 

Was there any specific challenge that was particularly hard to solve? How did you overcome it?

Ah, not exactly I would say. Some things have been harder than others, but most everything was solvable with some elbow grease. Including the Real Helpdesk tickets is a challenge which has yet to be solved, as there is currently no way to do so without heavy manual labor in data cleaning, which has been decided is not the best use of time. 

What have you learned from working on this project and collaborating with the ESA and Telespazio teams?

I can tell you about all the technical things I learnt, setting up LLM Agents, Figuring out RAG Chunking strategies, Model Fine-tuning, the list could go on forever. But to be quite honest the biggest takeaway I have had so far is an improvement in my ability to simplify technical concepts. My supervisors at ESA, while experienced in software, are not Machine Learning experts, and everything I did I had to explain to them in plain English, in understandable ways. While this is considerably easier to do so to a software engineer than the average person, I think I’ve gotten a lot better at this, and this has really helped me shine in conferences and presentations. At ADASS 2025 I had 12 minutes to present the entire project including a demo to a panel of astronomers/software engineers that were almost completely unfamiliar with the inner workings of LLMs. I think it went over really well and people seemed to really be able to grasp what the project consists of, and a great part of that comes from what I described above. 

How was your experience presenting your work at conferences like ADASS and AstroORDAS? How did the scientific community react to your work?

It was lovely. I’ve always enjoyed presenting and I was quite excited to show off the project to the greater community. They reacted very positively for the most part, the questions they asked after my oral presentations were exactly the kind of questions I was hoping for. It seemed to instill a sense of curiosity. One of the questions asked was whether I saw a future in projects such as this being expanded to multiple missions, or perhaps even ESA wide chatbots, which has always been the idea I have had in the back of my head. 

Do you feel your project opens new possibilities for integrating AI into space science workflows?

I would hope so. A lot of AI Agents in Space Science seem to be focused on solving a problem, i.e. using LLMs for some classification problem or for a specific bottleneck in some workflow. The project here is slightly different in the sense that it is built to be a general use agent, not directly focusing on a single problem, but rather attempting to be a generally useful tool for Space Science, making researchers lives a little easier. I hope that we can repeat this sort of project for other missions, based on the infrastructure that we have set up thus far.

With the new agreement signed between ESA and IE University, what kinds of opportunities do you envision for current and future students?

I am not exactly sure what the agreement entails, however I would hope that this would mean there is an influx of the younger generation moving toward ESA. While IE does not offer a Physics or Astronomy focused program (I think?), a lot of software engineers work at ESAC in Madrid, and I think we could always use some more. Especially with the rise of AI these days I would expect a growing number of opportunities to present themselves for IE students wanting to work with ESA. Now, I can only speak from the perspective of the ESAC in Madrid, but other standpoints across Europe likely have slightly different opportunities for students such as Robotics or R&D. 

How would you like to continue growing within the space sector?

Ideally I would like to expand this project into a multi mission idea, but for that I would likely need a team of people working on it. I suppose that could mean growing into a more technical lead–type role. That said, I don’t have any fixed ambitions at the moment.

What advice would you give to an IE student who hopes to work on real-world, high-impact projects?

Say yes to opportunities that present themselves. Your life at IE is filled with opportunities to work on real-world projects. Participate in Datathons, ask your professor whether they have anything you could work on with them, Check whether Final Projects proposals are offered from respectable companies. Just say yes.