Posters & Démonstrations
Découvrez tous les posters et démonstrations interactives présentés lors des AI Days 2026.
Démonstrations interactives
Renku: an open source platform for collaborative AI projects
Elisabet Capon, Rok Roskar
Swiss Data Science Center
AI projects combine code and data from diverse locations, such as git repositories, cloud storage platforms, specialized data repositories, and more. Coordinating these resources requires technical knowledge and represents time-consuming overhead that diverts time from core analytical work. Even more critically, this fragmentation presents a barrier to collaboration. When resources are not brought together in a repeatable, structured manner, team members struggle to replicate software environments or access necessary datasets, ultimately slowing project progress and hindering knowledge sharing. Renku addresses these challenges by providing an integrated platform that connects data, code, compute in a unified workspace. The platform enables scientists to launch browser-based interactive computing sessions with software dependencies pre-configured and data and code automatically accessible, thus eliminating setup friction. This approach transforms ad-hoc resource coordination into structured, reproducible projects that facilitate seamless collaboration. Built by the Swiss Data Science Center, Renku is an open source collaboration platform used nationwide for AI projects, research, and teaching. By streamlining the technical infrastructure, Renku allows teams to focus on analytical insights rather than resource management, empowering collaborative knowledge building.
Une IA souveraine et éco-conçue pour les PME suisses
Viet Hang Nguyen, Jean-Damien Beau
DeepAlps, e-Durable
L'intelligence artificielle (IA) s'est progressivement intégrée dans les habitudes des PME suisses, devenant un outil efficace et parfois indispensable au quotidien. Les employé(e)s recourent de plus en plus à des robots conversationnels pour diverses tâches, mais cette adoption soulève des défis majeurs en matière de protection des données sensibles et d'impact environnemental. En effet, les outils d'IA grand public tels que ChatGPT, Gemmini ou Claude ne répondent pas aux exigences strictes de protection des données et négligent souvent la question de l'empreinte écologique. Pour pallier ces lacunes, DeepAlps et e-Durable ont développé une solution d'IA éco-conçue, spécialement adaptée aux besoins des PME suisses. Notre solution se distingue par l'utilisation de modèles de taille raisonnable, de haute qualité et open source, garantissant ainsi une grande adaptabilité aux besoins spécifiques de chaque entreprise. De plus, l'hébergement repose sur le Recycled Cloud(R), une infrastructure entièrement suisse, optimisée pour minimiser l'impact environnemental. Ce projet illustre concrètement comment l'IA peut être intégrée de manière responsable dans les organisations régionales, en respectant des critères rigoureux de sobriété numérique, de durabilité et de confidentialité. Nous présentons une démonstration pratique de notre solution lors de cet événement, mettant en lumière son potentiel pour transformer les PME suisses tout en préservant les données sensibles et l'environnement.
Seamless ML deployment on high-end FPGAs for ultra fast inference using the hls4ml VitisAccelerator backend
Gaspard Le Gouic, Andres Upegui Posada, Quentin Berthet
HEPIA
This demonstration presents the XRT based VitisAccelerator hls4ml backend, a set of tools enabling the seamless deployment of complex ML algorithms on high-end FPGAs for ultra fast inference. A concrete exemple of a Convolutional Neural Network (CNN) trained on MNIST will be shown running on an FPGA accelerator using hls4ml and executed on a Versal VCK5000 platform. The model is converted into FPGA firmware using the VitisAccelerator backend, specifically developed at HEPIA and UNIGE to streamline integration with AMD/Xilinx platforms. The design incorporates quantization and pruning techniques to reduce computational complexity and hardware resource usage while maintaining classification performance. These optimizations enable efficient mapping of the neural network onto programmable logic with explicit control over parallelism and arithmetic precision. The live demonstration consists of executing inference workloads directly on the VCK5000 and measuring achieved throughput. By running the CNN in hardware and collecting performance metrics, we highlight the practical capabilities of FPGA based AI acceleration in terms of sustained throughput. Rather than focusing solely on theoretical advantages, the demonstration provides concrete performance observations obtained on real hardware. Originally inspired by low latency machine learning applications in particle physics at CERN, this approach illustrates transferable expertise in hardware aware AI deployment. While high energy physics remains a leading field for such developments, similar methodologies can be extended to industrial settings, for example in real time fault detection using autoencoder based anomaly detection. This work in progress showcases practical competency in FPGA based neural network deployment and highlights the strategic advantages of reconfigurable hardware compared to CPU and GPU based inference when low power, low latency and high throughput execution is required.
Evaluating Model Context Protocol Tool Use in Self-hosted LLMs
Lucie Juzan, Leonardo Angelini, Sébastien Rumley
HEIA-FR
The Model Context Protocol (MCP) has recently emerged as a standardized interface that enables Large Language Models (LLMs) to dynamically discover and invoke external tools. While MCP promises greater flexibility and improved generalization, its practical robustness and reliability across different models and operating conditions have not yet been systematically assessed. In this context, any development based on MCP should be accompanied by an ad hoc testing framework to evaluate its quality and usability in real-world settings. To address this need, we propose a methodology for designing such ad hoc evaluation frameworks. Our poster presents this methodology along with the architecture of the proposed testing framework, and includes comparative results across models and testing scenarios. Alongside the poster, we will showcase a live demonstration illustrating how MCP can be used to interact with a cellular automata simulator. The demo highlights that it is indeed possible to “talk to a simulator” using MCP. However, the poster emphasizes that possibility does not guarantee reliability: organizations aiming to deploy MCP in production must rigorously assess its reliability under realistic conditions. We will also demonstrate how the aforementioned methodology has been applied in the context of this demo to evaluate its performance and robustness.
Core AI Engine: a sovereign fast prototyping AI-On-Demand platform
Valentin Biolley, Simon Braillard, Andrea Petrucci, Ludovic Delafontaine, Bertil Chapuis, Jean Hennebert
HEIA-FR, HEIG-VD, Swiss AI Center
The Swiss AI Center Core Engine is a modular, open-source platform designed to bridge the gap between advanced research and practical industrial application for Swiss SMEs. This demo showcases the engine's ability to orchestrate complex AI workflows by chaining diverse services, from computer vision to specialized LLM pipelines, into seamless, production-ready solutions. Attendees will explore how the platform's "service-and-pipeline" architecture simplifies prototyping, allowing for the rapid deployment of trustworthy, Swiss-hosted AI models. By highlighting its Python-based backend and its intuitive frontend, the session demonstrates how businesses can achieve digital sovereignty and operational efficiency.
Posters
The ADVANCE toolkit: Automated descriptive video annotation in naturalistic child environments
Naomi Middelmann, Jennifer Glaus, Olga Sidiropoulou, Nastia Junod, Emily Wake, Jean-Paul Calbimonte, Manon Jaquerod, Kerstin Plessen, Matthew Vowels, Micah Murray
CHUV, UNIL, HES-SO
Video recordings are commonplace for observing human and animal behaviours, including interindividual interactions. In studies of humans, analyses for clinical applications remain particularly cumbersome, requiring human-based annotation that is time-consuming, bias-prone, and cost-ineffective. Attempts to use machine learning to address these limitations still oftentimes require highly standardised environments, scripted scenarios, and forward-facing individuals. Here, we provide the ADVANCE toolkit, an automated video annotation pipeline. The versatility of ADVANCE is demonstrated with schoolchildren and adults in an unscripted clinical setting within an art classroom environment that included 2–5 individuals, dynamic occlusions, and large variations in actions. We accurately detected each individual, tracked them simultaneously throughout the duration of the recording (including when an individual left and re-entered the field of view), estimated the position of their skeletal joints, and labelled their poses. By resolving challenges of manual annotation, we radically enhance the ability to extract information from video recordings across different scenarios and settings. This toolkit reduces clinical workload and enhances the ethological validity of video-based assessments, offering scalable solutions for behaviour analyses in naturalistic contexts.
Analyze and Optimize pipeline for Cherenkov Telescope SST-1M
Hugo Varenne, Andres Upegui Posada, Laurent Gantel, Jakub Kvapil, Matthieu Heller
HES-SO Master, HES-GE, UNIGE
Gamma-ray astronomy probes the highest-energy phenomena in the Universe through the detection of Cherenkov light produced by atmospheric air showers initiated by gamma rays. Imaging Atmospheric Cherenkov Telescopes (IACTs) use this light to reconstruct the type, energy, and arrival direction of the primary particle, a task made challenging by complex data and strong hadronic background contamination. This thesis investigates the application of Machine Learning, Deep Learning, and MLOps techniques to simulated data from the SST-1M (Small-Sized Telescope) of the Cherenkov Telescope Array Observatory, using the CTLearn library as the core framework. The objectives are to establish a production-ready environment for deep learning model generation and to enhance the model development workflow by integrating software engineering and MLOps best practices. SST-1M data are processed and configured to meet CTLearn requirements and high-performance computing constraints, and several deep learning architectures from the CTLearn library, including CNNs and ResNet models, are evaluated. The primary contribution is methodological rather than performance-driven. An MLOps-oriented framework is introduced that enables configuration-driven experimentation, automated report generation, and systematic model comparison. These additions improve reproducibility, traceability, and scalability, while model optimization techniques such as pruning and quantization are explored to reduce computational costs. This work demonstrates the suitability of MLOps for gamma-ray event reconstruction and provides a foundation for future CTLearn-based analyses at the intersection of astrophysics and modern machine learning engineering.
PolyPilot: An Agentic AI Platform for Safety-Critical Engineering (Swiss MedTech First)
Moad Kissai
PolyPilot
PolyPilot is an agentic AI platform (“Engineering OS”) for safety-critical engineering across regulated industries (MedTech, automotive, avionics, energy). It sits on top of existing toolchains (ALM/QMS/PLM, documents, spreadsheets) with a no-rip-and-replace approach. The core problem is not only building compliant systems, but continuously proving and updating compliance evidence as requirements and artefacts change across silos. Our approach combines hybrid retrieval (knowledge graph + semantic search) with deterministic, replayable agentic workflows and provenance-first outputs: every suggestion is source-linked, every action is logged, and humans remain in control through explicit approval gates. This enables repeatable workflows for cross-tool traceability maintenance, change impact analysis, and evidence updates—reducing ad-hoc document work and audit risk. We start with a Swiss MedTech wedge where MDR/IVDR evidence churn makes the ROI immediate. In the demo scenario, we run a GSPR mapping update after a requirement change and report time-to-update (from change detection to reviewer-ready export) as the primary metric, with human acceptance as the quality gate. Attendees will leave with a practical blueprint for deploying agentic workflows that are fast and defensible in audits.
SmartQ4S : des agents digitaux pour une assistance intelligente et durable dans les PME industrielles
David Rossy, Nicolas Ramosaj, Beat Wolf, Jean Hennebert, Régis Vonarb
HEIA-FR
L’essor de l’Industrial AI et des modèles de langage de grande taille (LLM) ouvre de nouvelles perspectives pour la supervision intelligente des systèmes de production. Toutefois, l’intégration conjointe des données qualité, des pratiques d’amélioration continue et des exigences de durabilité demeure un défi majeur, en particulier pour les PME industrielles. Le projet Smart Quality for Sustainability (Smart Q4S) explore comment des agents digitaux peuvent soutenir la résolution de problèmes qualité tout en contribuant à l’amélioration des indicateurs environnementaux. La question de recherche porte sur la capacité d’architectures hybrides LLM-RAG (Retrieval Augmented Generation) à s’adapter aux contraintes industrielles : exigences en matière de sécurité et de confidentialité, bruit et variabilité des données opérationnelles, rareté des anomalies et nécessité d’intégrer le savoir‑faire implicite de la production. La méthodologie repose sur une recherche appliquée conduite avec un consortium industriel, combinant analyse de processus, ingénierie des données et développement de prototypes d’agents digitaux exploitant des bases de connaissances textuelles (procédures, historiques de défauts, standards qualité et durabilité) complétées par les données de productions. Les travaux présentés correspondent à des résultats préliminaires issus de prototypes en cours de validation. Ils mettent en évidence le potentiel des agents LLM-RAG pour améliorer l’aide à la décision, réduire le temps de diagnostic des défaillances et quantifier les impacts environnementaux associés aux dérives de processus. Les premiers résultats relèvent des défis techniques liés à l’explicabilité et au réglage des hyperparamètres (vectorisation, prompts, modèles). L’impact attendu de Smart Q4S réside dans le développement d’agent digitaux industriels pragmatiques, explicables et transférables, capables de relier supervision qualité, performance économique et durabilité.
On-device AI: Building a Custom Speech-to-Speech Pipeline for Rare Languages
Timothée Van Hove, Rémy Marquis, Bertil Chappuis
HEIG-VD
Hundreds of languages are spoken by millions of people but remain invisible to mainstream speech technology. System-provided speech APIs cover 60 languages; the remaining hundreds are simply absent. Cloud-based solutions can partly help, but they break two hard requirements: privacy and offline access. State-of-the-art speech-to-speech models like Hibiki show impressive end-to-end quality, but they don’t yet cover our target languages. We built a custom three-stage pipeline to answer a concrete question: Can a 4-year-old phone run a full translation pipeline, supporting 100+ languages entirely on-device?
KOMKI: Intelligent Email Routing with LLMs
Mennan Selimi, Hamit Kamberi, Meral Selimi
SEE University
Organizations struggle with inefficient and error-prone email handling. Manual processing takes 2–5 minutes per email, and misclassification rates can reach 30%, causing delays, customer dissatisfaction, and increased workload. Traditional rule-based and classical machine learning approaches often fail in multi-label, multilingual, and ambiguous scenarios common in real-world communication. KOMKI proposes a three-phase intelligent email routing framework combining semantic keyword extraction, context-aware classification using large language models (GPT-3.5, GPT-4, LLaMA), and automated routing with summarization. The system supports multi-label classification, multilingual input, and includes a human-in-the-loop option for sensitive cases. Evaluated on real organizational datasets and benchmarked against traditional methods, KOMKI demonstrates improved accuracy, scalability, and efficiency. A cost-performance analysis comparing cloud APIs and self-hosted models further provides practical guidance for deployment.
Solving Finite Element Simulation Models Using Physics-Informed Neural Networks
Antoine Ottiger, Jean-Luc Robyr, Beat Wolf
HES-SO
The finite element method (FEM) is widely used in mechanical engineering to perform accurate physics-based simulations. Nevertheless, it remains computationally expensive, which can be a limitation depending on the application. This project aims to use neural networks (NNs) as a substitute for FEM simulations in the context of modal analysis. We compare the performance of two loss functions: a fully data-driven approach based on the mean squared error (MSE) and a physics-informed loss derived from FEM formulations. Within the scope of our experiments, we found that the MSE loss was more suitable, providing more accurate results with significantly lower training computation time. We then exploit the trained NN model in a real-time system, enabling the estimation of physical parameters in a few seconds, compared to several hours when using FEM simulations.
Boosting innovation process using LLM agents
Hatem Ghorbel, Constant Ondo, Simon Fuhlhaber, Lamia Ben Hamida, Stefanie Hasler, Henrique Marques Reis, Brendan Studer, Guy-Raphaël Stauffer
HE-ARC ING, PICC, HE-ARC GES
PICC Software SA aims to empower teams to share experiences, situations and ideas to seek help from, or support, their community in solving problems. PICC proposes using the TRIZ method to help companies develop innovative solutions. The TRIZ method follows a multi-step process that transforms a problematic contradiction into an innovative solution. The goal of this project is to use LLM-based agents to support each step of this process. Humans remain in the loop, validating the output of the LLM at every stage to detect potential deviations and ensure the process stays on track. Human oversight also helps prevent the propagation of hallucinations across subsequent steps.
AD-DITION – Writing Style Cloning
Jean Hennebert, Célien Donzé, Eden Brenot, Jérémy Marchon
iCoSys HEIA-FR
The AD-DITION project addresses the limitations of generic LLM outputs by developing a personalized generative engine in order to reproduce the writing style of a person. Our methodology relies on a curated dataset of a user’s professional email correspondence used for fine-tuning a model, specifically targeting stylistic markers and vocabulary. In this poster, we will show that fine-tuning an LLM using LoRa to imitate style gives better results than only giving examples of the writing style of a person. Specific metrics used for style evaluation will be presented as well as an approach using an LLM-as-a-Judge.
HydroScan: AI-powered real-time water-course surveillance
Marc Lany, Olivier Chabloz, Vincent Roch, David Lavanchy, Tristan Brauchli, Snežana Nektarijević, Kyle van de Langemheen, Isione Bonvalot, Ivan-Daniel Sievering
AIMsight, HEIG-VD, SDSC
Hydroscan is an innovative, non-invasive flow monitoring solution powered by stereoscopic vision and advanced artificial intelligence. Traditional methods for measuring flow in natural rivers and artificial channels often rely on invasive infrastructure such as artificial weirs. These structures disrupt ecosystems and frequently deliver unreliable data during extreme events like floods. Hydroscan replaces these limitations with a vision-based approach. Using advanced AI-driven computer vision algorithms, the system reconstructs the 3D geometry of the riverbed or channel, measures water levels and surface velocities, and accurately calculates discharge — all without altering the natural environment. Designed for real-world conditions, Hydroscan adapts to variable lighting, complex river morphologies, and dynamic flow conditions. AI-based segmentation techniques enhance detection robustness and significantly reduce sensitivity to environmental disturbances. The system is developed by AIMsight in collaboration with HEIG-VD, which contributes expertise in computer vision algorithms and hydraulic testing, and the SDSC, which actively supports the development of AI methodologies. Hydroscan enables more reliable and sustainable water resource management. It addresses critical challenges including sediment monitoring, flood risk management, hydropower operations optimization, and ecosystem protection — providing a scalable and environmentally responsible alternative to conventional monitoring systems.
Predicting Pathogen-Specific Phage Lysis: A Scalable AI Framework for Precision Medicine
Farzaneh Labbaf, Wan-Ting Huang
Precise Health
Problem Statement: As antimicrobial resistance (AMR) outpaces traditional antibiotic development, bacteriophage therapy has emerged as a critical alternative for resistant infections. However, the high specificity of phages requires a precise matching process that currently relies on time-consuming laboratory culture, often taking weeks. This delay is a primary barrier to treating acute and sub-acute infections where the window for intervention is measured in hours. Methodology and results: We present the Digital Phagogram, an AI-driven platform that automates the identification of effective phage-bacteria pairings through genomic analysis. Our approach utilizes an ensemble modeling framework trained on large-scale phage-bacteria interaction datasets, integrating both public repositories and in-house experimental data. To ensure clinical utility, the system provides two distinct metrics for decision support: a predicted lytic probability and a confidence score based on the genomic similarity between the query samples and the training manifold. The model is specifically designed to generalize to novel pathogens and phages, a necessity for real-world clinical deployment where unique bacterial strains frequently emerge. The system was validated using Leave-One-Bacteria-Out (LOBO) cross-validation to rigorously test its predictive power on completely unseen bacterial hosts. At an 80% lysis probability threshold, the model achieved a Hit@1 of 85%, a Hit@2 of 94%, and a Hit@3 of 98%. These results demonstrate that the ensemble architecture effectively captures the complex features of phage-host specificity, compressing the sample-to-treatment timeline from weeks to approximately 24-48 hours. Significance: By replacing stochastic wet-lab screening with a deterministic computational framework, this platform offers a scalable solution for personalized medicine. The integration of confidence scoring provides clinicians with the necessary transparency to make informed treatment decisions, directly addressing the logistical and biological barriers of the AMR crisis.
HR Professionals and Artificial Intelligence: A (R)evolution of Roles
Justine Dima, Guillaume Revillod
HEIG-VD, ENAP
This study investigates how AI influences the traditional HR roles and the emerging capabilities required in AI-enabled organizations. The study adopts a qualitative research design based on 26 semi-structured interviews with HR professionals in Switzerland and two expert focus groups. Findings indicate that AI transforms all five HR roles but with different dynamics. Administrative and operational tasks are increasingly automated, allowing HR professionals to redirect their efforts toward strategic and developmental activities.
Mobile Image-Based Apple Detection for Yield Estimation in Orchards
Cédric Campos Carvalho, Elena Najdenovska, Ylli Fazlija, Fabien Dutoit, Laura Elena Raileanu
HEIG-VD
Accurate yield estimation is essential for optimizing orchard management and maximizing the proportion of high-quality fruits. However, current practices rely largely on manual fruit counting, which is time-consuming and often leads to large estimation errors. Existing automated approaches frequently depend on complex and costly sensing systems, limiting their adoption by growers. This project investigates a practical and accessible method for estimating apple production using smartphone images combined with fruit growth modelling. A mobile application was developed to capture images of apple trees and automatically upload them to a cloud-based processing platform for automated analysis. Apple detection is performed using a YOLOv11 object detection model, whose accuracy and speed make it particularly suitable for real-time smartphone-based applications. The model was trained on 826 annotated images collected from commercial orchards and public datasets, covering several apple varieties, fruit development stages, and acquisition conditions. Fruit counts are derived from detection results using a regression approach. Detection results are combined with a previously established fruit growth model based on seasonal fruit diameter evolution, enabling fruit volume to be estimated from the detected fruit count and the predicted fruit diameter at the time of image acquisition. The detection model achieved a mean Average Precision (mAP) of 87.5% for detections being considered correct at an Intersection over Union threshold of 50%. Fruit counting showed strong agreement with annotated data (R² = 0.95), with a mean absolute error of approximately 7 fruits per image. This work demonstrates the feasibility of an automated low-cost yield estimation based on images taken with a smartphone, thereby representing a promising step toward scalable digital tools for precision fruit production. Project funded by the HES-SO, grant number 130502/IA-RECHERCHE23-42.
Quantifying Deforestation Risks in Palm Oil Production
Firas Dridi, Dimitri Kohler, Antoine Lestrade
HE-ARC
The objective of this project is to provide foundations and decision-makers with a scientifically rigorous tool to identify, measure, and visualize the risk of deforestation within the industrial palm-oil supply chain. Current approaches often rely on static maps that show where mills are located but fail to explain why deforestation occurs. This project moves beyond simple mapping to create a dynamic Digital Twin of the supply chain. By integrating geospatial data with economic logic, we build a model that assesses the “appetite” of industrial actors to consume illegal or unverified palm oil.
Autonomous Indoor Quadcopter Navigation in GNSS-Denied Environments
Diego Fraile, Denis Rosset, Roland Scherwey, Jean Hennebert
Groupe de Recherche Interdisciplinaire Drones, HEIA-FR iCoSys, HEIA-FR iSIS
Indoor autonomous navigation of unmanned aerial vehicles is challenging because Global Navigation Satellite System signals are unavailable in most indoor environments. This work presents a navigation framework that enables a quadcopter to operate autonomously in GNSS-denied spaces. The system employs visual SLAM using a stereo camera, with processing performed on an NVIDIA Jetson Orin Nano companion computer. Ground-relative velocity measurements from an optical flow sensor are fused with SLAM pose estimates and inertial data through an Extended Kalman Filter implemented in the PX4 flight controller, producing a robust state estimate for closed-loop control. Obstacle avoidance is achieved by combining the SLAM-generated map with depth information from the stereo camera within the Nav2 navigation stack, which computes safe two-dimensional trajectories while maintaining a fixed flight altitude. The framework was validated both in Gazebo simulation and through real-world flight experiments conducted in a warehouse environment. In these tests, the quadcopter successfully navigated around obstacles and reached a commanded goal pose autonomously.
A Dynamic Proxy-Based Framework for Clustering in Distributed Learning
Mohsen Salimi Khanghah, Khalid Ali, Phil Aupke, Andreas Kassler, Nabil Abdennadher, Giovanna Di Marzo Serugendo
University of Geneva, Deggendorf Institute of Technology, Karlstad University, HEPIA
Distributed learning paradigms, such as Federated Learning (FL) and Split Learning (SL), enable privacy-preserving analytics across decentralized clients; however, their performance often deteriorates under the significant statistical heterogeneity common in real-world data, which violates the non-IID assumptions of standard algorithms. Multi-model specialization through client clustering offers a potential remedy to this data heterogeneity; nevertheless, traditional approaches treat clustering as a static pre-processing step decoupled from the learning dynamics. In practice, specialization is a dynamic process in which client updates, proxy representations, and aggregation topologies evolve together. This paper introduces a dynamic framework for FL and SL, in which the client’s proxy serves as the feedback variable that represents the client’s learning characteristics. The proposed framework integrates hierarchical clustering with geometry-aware routing, utilizing distances derived from information geometry to determine model assignments. We evaluate the framework using a real-world electricity forecasting task under both FL and SL architectures. Our results demonstrate that the multi-model specialization improves global Mean Squared Error (MSE) by up to 13% and improves tail robustness by more than 14% relative to single-model baselines. We identified a performance boundary where aggressive reassignment and automatic cluster selection induce topology collapse. According to the results, we are providing practical design guidelines for implementing reliable, distributed learning models in heterogeneous environments.