I am a Machine Learning Researcher. I currently work as a Doctoral Researcher at ETH Zurich, focusing on advancing healthcare with ML-based Digital Biomarkers. My research interests lie at the intersection of AI/ML, medicine, mobile applications and ubiquitous computing.
Having discovered a passion for programming and engineering at age 11, I have acquired diverse skills in various programming languages, systems, and frameworks over the years. I completed multiple projects and applications in the fields of Machine Learning, Computer Vision, Healthcare and Embedded Systems.
I am one of the winners and national finalists of Jugend forscht in 2016. Further, I support the lecture Developing Digital Biomarkers as a teaching assistant, and have supervised around 20 students in their semester, bachelor and master theses.
PhD in Machine Learning & Applied Computer Science, 2025 (expected)
ETH Zurich - Swiss Federal Institute of Technology
MSc in Computer Science - Machine Learning & Medical Technologies, 2021
Ilmenau University of Technology
BSc in Computer Science, 2020
Ilmenau University of Technology
I am currently a third year PhD researcher, focusing on machine learning in healthcare as well as on mobile- and edge devices. My research topics include:
Additionally, I am a teaching assistant for the lecture Developing Digital Biomarkers.
I completed various projects for different clients worldwide, especially in the fields of:
List of clients and projects under NDA.
As a research engineer, I worked on machine learning applications in medicine (especially Parkinson), and embedded applications for space- and high radiation environments.
As a research engineer, I worked on medical sensors and medical applications.
Optimized a proprietary FPGA hardware accelerator implementation for Convolutional Neural Networks and computer vision algorithms for person detection and -segmentation from RGB videos in the context of autonomous driving on a SoC platform.
Worked on indoor robotic applications in the context of patient-assistance, rehabilitation and elderly care.
Explainable deep-learning time series extrinsic regression methods based on counterfactual approaches and evolutionary algorithms. The goal is to generate recommendations for improving general health using time-series models (TCNs) for biological age estimation.
Development of a Health Asistant for asthma. Using 5 digital sensors and AI methods to passively assess track symptons and predict exacerbations in a 1-year-long study with 160 participants.
Longitudinal tracking of circadian rhythms of individuals using commercial devices in a 14-day study with 36 participants.
An open-source cross-platform, cross-language framework for machine-learning- and data collection applications across cloud, edge, mobile and wearable devices. Supports to run ML models directly in Python, even on Android and WearOS.
Development of a novel open source smart inhaler compatible with the three most common inhaler types (MDI, Turbuhaler and Diskus). Using passive sensors such as accelerometers to infer inhaler usage using machine learning.
Implementation of an outdoor Visual Scene Change Detection (VSCD) using deep learning computer vision models, outperforming state of the art approaches by pretraining on much larger indoor datasets (ChangeSim).
Implemention of realtime capable Human Activity Recognition (HAR) approach based on deep learning models SlowFast and PoseC3D. A major challenge was slow the inference speed of current state-of-the-art models. Several improvements have been made to achieve realtime inference (10+ FPS).
Modification of a popular video game to extract high-quality synthetic data for autonomous driving scenarios, such as road sign annotations, person pose estimation and object segmentation. Segmentations were extracted by hijacking the rendering pipelines of the game and extracting images at various stages of the rendering process.
Trained and implemented Temporal Convolutional Neural Networks (TCNs) for detecting Freezing of Gait. Deployed the neural networks on FPGAs using VitisAI, adhering to strict realtime requirements.
Implemented a hardware accelerator for Convolutional Neural Networks for person detection on FPGAs using HLS. Implemented highly optimized pre- and postprocesing of images using SIMD on a microcontroller. Outperformed previous results by 5x (1040ms to 195ms per frame).
Implementation of a realtime-capable 3D semantic mapping framework based on Normal Distribution Transform (NDT) and semantic histogram maps. Mapping leverages hardware acceleration on embedded devices using TensorRT.
A realtime road sign recognition App for smartphones incorporating object detection and classification (SSDLite, MobileNet), optionally using hardware acceleration (CoreML, TensorFlow). Currently standing at 5000+ downloads.
Designed and built (PCB, soldering, programming) from the ground up a device capable to record and analyze biosignals via electromyografie using electrodes. Allows to play games and write texts. Presented at the Jugend forscht national finals 2016.
In this paper, we present CLAID, an open-source cross-platform middleware framework based on transparent computing compatible with Android, iOS, WearOS, Linux, macOS, and Windows. CLAID enables logical integration of devices running different operating systems into an edge-cloud system, facilitating communication and offloading between them, with bindings available in different programming languages. We provide Modules for data collection from various sensors as well as for the deployment of machine-learning models. Furthermore, we propose a novel methodology, ML-Model in the Loop for verifying deployed machine learning models, which helps to analyze problems that may occur during the migration of models from cloud to edge devices. We verify our framework in three different experiments and achieve 100% sampling coverage for data collection across different sensors as well as an equal performance of a cough detection model deployed on both Android and iOS devices. Additionally, we compare the memory and battery consumption of our framework across the two mobile operating systems.
Circadian rhythms govern biological patterns that follow a 24-hour cycle. Dysfunctions in circadian rhythms can contribute to various health problems, such as sleep disorders. Current circadian rhythm assessment methods, often invasive or subjective, limit circadian rhythm monitoring to laboratories. Hence, this study aims to investigate scalable consumer-centric wearables for circadian rhythm monitoring outside traditional laboratories. In a two-week longitudinal study conducted in real-world settings, 36 participants wore an Actigraph, a smartwatch, and a core body temperature sensor to collect activity, temperature, and heart rate data. We evaluated circadian rhythms calculated from commercial wearables by comparing them with circadian rhythm reference measures, i.e., Actigraph activities and chronotype questionnaire scores. The circadian rhythm metric acrophases, determined from commercial wearables using activity, heart rate, and temperature data, significantly correlated with the acrophase derived from Actigraph activities (r=0.96, r=0.87, r=0.79; all p<0.001) and chronotype questionnaire (r=-0.66, r=-0.73, r=-0.61; all p<0.001). The acrophases obtained concurrently from consumer sensors significantly predicted the chronotype (R2=0.64; p<0.001). Our study validates commercial sensors for circadian rhythm assessment, highlighting their potential to support maintaining healthy rhythms and provide scalable and timely health monitoring in real-life scenarios.
Encouraging people to manage their health is essential in preventing chronic diseases like type 2 diabetes and heart disease. With mobile technologies such as health apps and fitness trackers, monitoring health has become more accessible and affordable compared to doctoral visits and traditional health check-ups. These technologies are, however, mainly used for disease management rather than prevention and there is no common understanding of how they can best be used for preventive purposes. To this end, we introduce the Bitemporal Lens Model a comprehensive method to use mobile technologies for disease prevention. We explain the structure and usefulness of the Bitemporal Lens Model, discuss its advantages and limitations, and present potential use cases.
A key proficiency an autonomous mobile robot must have to perform high-level tasks is a strong understanding of its environment. This involves information about what types of objects are present, where they are, what their spatial extend is, and how they can be reached, i.e., information about free space is also crucial. Semantic maps are a powerful instrument providing such information. However, applying semantic segmentation and building 3D maps with high spatial resolution is challenging given limited resources on mobile robots. In this paper, we incorporate semantic information into efficient occupancy normal distribution transform (NDT) maps to enable real-time semantic mapping on mobile robots. On the publicly available dataset Hypersim, we show that, due to their sub-voxel accuracy, semantic NDT maps are superior to other approaches. We compare them to the recent state-of-the-art approach based on voxels and semantic Bayesian spatial kernel inference~(S-BKI) and to an optimized version of it derived in this paper. The proposed semantic NDT maps can represent semantics to the same level of detail, while mapping is 2.7 to 17.5 times faster. For the same grid resolution, they perform significantly better, while mapping is up to more than 5 times faster. Finally, we prove the real-world applicability of semantic NDT maps with qualitative results in a domestic application.
In this paper we report on our implementation of a temporal convolutional network trained to detect Freezing of Gait on an FPGA. In order to be able to compare our results with state of the art solutions we used the well-known open dataset Daphnet. Our most important findings are even though we used a tool to map the trained model to the FPGA we can detect FoG in less than a millisecond which will give us sufficient time to trigger cueing and by that prevent the patient from falling. In addition, the average sensitivity achieved by our implementation is comparable to solutions running on high end devices.