At Zalando, I am part of the Zalando Assistant team. I built the core ML components of the multi-agent AI layer powering Zalando Assistant; LLM-based router, memory-augmented personalisation, millisecond-latency intent classifier, and malicious-intent detector with drift monitoring. These components serve millions of customer interactions across 20+ European markets. Previously, as part of the Algorithmic Privacy and Fairness team, I built a toolbox for robustness, privacy and fairness assessments of ML models, and developed novel adversarial ML techniques and privacy attacks to assess the robustness of ML systems.

I drive applications of LLMs and multimodal LLMs in semantic search and information retrieval. I designed an AI intervention mechanism for the Assistant's integration into Zalando's search engine, significantly increasing traffic share and customer conversions. I also built a large-scale evaluation framework [ECIR 2025] [blog] matching human quality at ~1000x lower cost, now adopted org-wide across teams.

I lead AI safety R&D for Zalando Assistant, including red-teaming, risk assessment, and building a malicious-intent classifier with drift monitoring on live traffic (The Wall Street Journal, 2024).

I am also a collaborator and advisor to the Earth Rover Program (The Guardian, 2025), contributing on ML, collaborative software development, and engineering best practices.

At the Alan Turing Institute, I built MapReader, an end-to-end computer-vision pipeline for classifying ~16k historical maps (~30M geospatial patches) at scale [SIGSPATIAL] (winner of the Roy Rosenzweig Prize), created DeezyMatch for fuzzy string matching [EMNLP], trained four neural language models on ~5B tokens of historical English [JOHD], and worked on privacy-preserving ML including differential privacy [arXiv], generative models, and federated learning, RL for air-traffic control (Project Bluebird), dynamic spatial microsimulation (Daedalus), contributed to Scivision (a toolkit for scientific image analysis), and OCR quality for NLP [ICAART].

At the University of Oxford, I designed computational imaging pipelines and inverse-problem solvers for the ERC-funded DEEP TIME project, fusing heterogeneous seismic observations into 3-D models of Earth's deep interior (GJI; Nature Geoscience; Quanta Magazine). I created SubMachine, web-based tools for exploring 3-D geophysical models (750k+ queries served), and built obspyDMT, an open-source Python toolbox for retrieving and processing large seismological datasets. During my PhD at LMU Munich, I developed large-scale computational methods for seismic waveform tomography on supercomputers, including the NDLB (No Data Left Behind) algorithm for multifrequency body-wave tomography.

49 peer-reviewed publications, ~85 conference contributions, >2,500 citations (h-index 25; Google Scholar). Also author of two book reviews and four published technical reports.