🔬Machine Learning Engineer & AI Researcher at DFKI
I am a Machine Learning Engineer and AI Researcher at DFKI in Kaiserslautern, building systems where research credibility and production reliability both matter.
My current work spans three applied domains: multimodal crop-yield prediction combining satellite imagery, weather data, and geospatial context; privacy auditing for ML models and LLMs including membership inference and model inversion risk; and MLOps infrastructure that turns research into repeatable, deployable pipelines.
That work has contributed to peer-reviewed publications, including CVPR 2026, Remote Sensing of Environment, Computers and Electronics in Agriculture, and IEEE venues. It has also produced deployed pipelines with industry partners in Germany, which keeps my work grounded in operational constraints rather than research prototypes alone.
My path here was non-linear. I started with electronics engineering at HBTI Kanpur, then spent four years at IBM building distributed Java EE systems for enterprise telecom before moving into AI through an M.Sc. in Cognitive Science at Osnabrück University. That engineering background still shapes how I think about ML: the methods should be sound, but the systems also need to run, scale, and survive real-world use.
Multimodal crop-yield prediction
Field and sub-field scale yield forecasting using Sentinel-2 imagery, weather data, soil information, and harvester GPS records. The focus is not only model performance, but also building data and training pipelines that hold up across real farming workflows. The resulting systems reached R² up to 0.83 at field resolution, supported deployments with industry partners in Germany, and produced publications in CVPR 2026, Remote Sensing of Environment, and Computers and Electronics in Agriculture.
Privacy auditing and trustworthy AI
Practical privacy risk assessment for ML models and LLMs, covering membership inference, model inversion, and compliance workflows under EU AI Act requirements for healthcare applications. My goal here is to turn privacy analysis into a repeatable engineering workflow with measurable testing, reporting, and decision support rather than a box-checking exercise.
Production ML systems
Turning research code into reliable, reproducible infrastructure with Kedro pipelines, MLflow experiment tracking, Apache Airflow orchestration, Docker, and CI/CD. This thread cuts across all my projects and reflects the part of my work I value most: making strong ML ideas usable outside notebooks.
Recent papers connected to my work in remote sensing, multimodal learning, and trustworthy AI.