Welcome to my homepage! I'm Pascal Sager, a PhD candidate at the Centre for Artificial Intelligence (CAI) at the Zurich University of Applied Sciences (ZHAW), a visiting doctoral student at the Institute of Neuroinformatics at the University of Zurich and ETH Zurich, and an Associated PhD Student at the ETH AI Center. My research is supervised by Prof. Dr. Thilo Stadelmann (Director of CAI) and Prof. Dr. Benjamin Grewe (Head of the Neural Learning and Intelligent Systems Lab at ETH).
My research focuses on advancing AI systems by incorporating general knowledge about the world. Drawing inspiration from neuroscience, I explore how deep learning systems can benefit from biologically motivated principles, particularly in how they form and structure internal representations. As part of this work, I investigate world models, trying to make current system more efficient and flexible. My research is supported by a personal fellowship from the Digitalization Initiative Zurich, Switzerland.
Alongside my research, I serve on the board of the Sustainable Impact Program at ZHAW, where I promote the ethical and responsible use of technology. At CAI, I also lead AI demonstrations and workshops, helping to make AI accessible and engaging for a broad range of audiences.
I am available for talks and presentations on AI and learning systems. I have experience with a range of formats, including corporate sessions (e.g., Roche), interactive workshops, exhibitions (e.g., Night of Technology, Scientifica), and public talks (e.g., Digital Night Winterthur).
Thank you for visiting my page. I look forward to sharing my research and perspectives with you.
Affiliated with the ETH AI Center as an Associated PhD Student, supporting interdisciplinary research at the interface of robotics and artificial intelligence.
Visiting PhD student in the Neural Learning and Intelligent Systems group led by Prof. Dr. Benjamin Grewe. My research explores biologically inspired learning mechanisms for building structured internal models in artificial agents.
I began at CAI as a research assistant in computer vision. After 18 months, I took over as head of GPU infrastructure and AI demonstrators. In October 2023, I started my PhD on advanced learning algorithms, shifting focus from infrastructure to research and the development of AI demonstrator systems. I also serve on the board of ZHAW’s Sustainable Impact Program, where I help shape initiatives around responsible technology and sustainability.
At AlpineAI, I contribute to the development of privacy-preserving AI technologies, focusing on SwissGPT and enterprise-level assistive agents. The company specializes in deploying Large Language Models (LLMs) with strict guarantees for data protection in regulated industries.
Before transitioning into AI, I worked in hardware and software engineering across multiple companies. My work included embedded systems, IoT applications, and full-stack software development. A detailed overview is available on my LinkedIn profile.
In this project, I explored the use of large language models (LLMs) to generate SQL queries from natural language questions. My approach focused on enhancing precision and generalization without relying on proprietary models or large-scale datasets. I surpassed the current state-of-the-art (SOTA) performance on all three ScienceBench datasets, demonstrating the capability of lean, carefully optimized LLM pipelines. Although the results are strong, this work has not yet been published.
View ProjectFor the CLEF 2025 competition, I tackled the task of retrieving relevant scientific papers based on social media posts. Competing against 31 international teams, I secured 3rd place overall. What makes this achievement stand out is that I relied solely on small, open-source models and free resources, avoiding paid APIs entirely. This project highlights the potential of resource-efficient retrieval methods in scientific domains.
Dedicated to fostering a culture of sustainability, I initiated and lead the project AIMS – AI Infrastructure Manager for Sustainability. This initiative, designed to address the environmental impact of our university's AI operations, embodies a holistic approach. We're implementing an energy consumption monitoring system, introducing an efficient job queue system through Slurm, and look for ways to ingeniously repurpose dissipated heat. These measures aren't merely elevating standards within the Centre for Artificial Intelligence (CAI); they're sparking a broader transformation across the entire ZHAW community, as we collectively champion responsible resource utilization and eco-conscious practices.
In my second project thesis during my master's program, I laid the foundations for machine learning-driven body composition analysis at the Cantonal Hospital of Aarau (KSA). This thesis not only led to a publication but also to a funded research project. The ML-BCA project is bringing the findings of my project thesis to life as a practical product for KSA. Our primary objective in this endeavor is to develop a robust application for facilitating medical validation and prospective studies, and to disseminate our collective scientific contributions to the broader community.
View ProjectIn my Master's thesis, I introduce an innovative learning framework by combining insights from neuroscience. While deep learning excels in automated image analysis, it faces issues like noise sensitivity, limited object recognition adaptability, and a constant need for extensive training data. In contrast, the human brain excels in holistic, non-linear image feature processing through self-organization and local learning. This thesis pioneers an image-processing approach inspired by the brain's operations. Empirical results, including Hebbian-trained lateral connections, demonstrate remarkable robustness and the ability to recover occluded objects.
View ProjectIn my role as Head of AI Demonstrators, my primary mission is to develop compelling showcases of AI's prowess, and a standout amongst these is our autonomous robodog. This remarkable creation undergoes continuous enhancement through collaboration with our talented students, who engage in project work and Bachelor's theses. Equipped with a camera and a LIDAR sensor, the robodog processes this data to navigate its surroundings safely. It's not just a marvel of technology; it's a responsive companion, capable of recognizing gestures, planning actions, and executing them with autonomy.
In my role as Head of AI Demonstrators, my primary responsibility involves developing cutting-edge AI showcases. Among these, we developed a remarkable Swiss-German speech-to-image generator, a multi-user web application. It effortlessly transcribes Swiss-German speech to text, refines the text into a polished prompt, and subsequently generates an image that corresponds to the prompt. Take a glance at the left, and you'll discover the result for the intriguing phrase, 'Älien Pflanze' which translates to 'alien plant'.
View ProjectAlpineAI pioneers the utilization of Language Model (LLM) technology to create exceptional value for businesses, prioritizing the utmost data security and privacy. I spearheaded the research and development of SwissGPT, a cutting-edge LLM designed to securely access sensitive corporate data, thereby unlocking the full potential of your company's knowledge through the power of AI.
View ProjectMonitoring diverse sensor signals of patients in intensive care can be key to detect potentially fatal emergencies. But in order to perform the monitoring automatically, the monitoring system has to know what is currently happening to the patient: if the patient is for example currently being moved by medical staff, this would explain a sudden peak in the heart rate and would thus not be a sign of an emergency. Therefore, the system is extended with video analysis capabilities and movements of the patient and the medical staff are detected.
View ProjectBody composition analysis can improve patient prognostication and contribute to higher safety, efficiency and quality of the patient journey. This analysis is often neglected due to the high manual effort or the lack of labels to develop a system for automatic processing. This project thesis proposes an automated method for body composition analysis that can be carried out on 3D computer tomographic (CT) scans without labels and with limited computational resources.
View ProjectIn a previous project, a solution to translate printed music scores into machine-readable music sheets was developed. However, it only works for high quality input. To scale up business, it should work as well for smartphone pictures, used sheets etc. Project RealScore enhances the successful predecessor project by making deep learning adapt to unseen data through unsupervised learning.
View ProjectThis thesis demonstrates the effectiveness of Gated Recurrent Units (GRU) in predicting speech frames up to a single word's length. Our model, trained on the TIMIT dataset, predicts subsequent frames of a Mel-spectrogram from a limited set of initial frames. We provide insights into the ideal number of input frames, frame-level, and phoneme-level prediction accuracy. As a pioneering work in this field, we emphasize the lessons learned and suggest potential enhancements for future research endeavors.
View ProjectFWA is a project funded by Innosuisse, that focuses on automating food waste classification. When food is discarded, an instant photo is captured, enabling us to calculate the difference from the previous state. We meticulously segment and estimate the weight of the discarded items. This data empowers kitchens to refine menu planning, thereby reducing food waste and enhancing sustainability.
View ProjectI successfully completed a Reinforcement Learning course at Zurich University of Applied Sciences, culminating in a challenging 3-day hackathon where I excelled and earned the top grade of 6.0. I've shared the code at https://github.com/sagerpascal/rl-bootcamp-hackathon, featuring highly efficient algorithms for Lunar Lander, yielding outstanding results in sample effectiveness, training duration, and average reward. Furthermore, I've meticulously documented the project for your reference.
View ProjectIn my bachelor thesis, the digitalization of the products of the company Spühl GmbH is examined on the basis of different aspects. The existing processes are critically scrutinized in order to minimize the project and product risk and to involve the customer more closely into the development process. The proposed concept for new processes is validated with the development of a new software product. The application enriches the already captured machine data with data from the production process. The application consists of a Java backend (with spring boot), a swagger interface and a React frontend. However, details cannot be provided due to a non-disclosure agreement. The Bachelor thesis was graded with the highest grade 6.0.
This was my first deep learning project in the field of NLP and is therefore featured on my homepage. According to a guide at (https://developers.google.com/machine-learning/guides/text-classification/step-2-5?hl=hi), I trained a sepCNN model on about 150,000 Twitter posts. In the end, the trained model predicted with an accuracy of 96% whether a given text is hate speech.
View ProjectIn the project thesis during my Bachelor's studies, I developed a solution for data acquisition of machines from the company Spühl GmbH together with a fellow student. The application is based on Azure IoT. Different docker containers collect various data, then this data is optimized with Stream-Analytics, sent to the cloud and optionally stored in a database or visualized using PowerBI.