Arthur Tabary
Description
Reef Pulse, founded in 2021 and based in Saint-Denis, Reunion Island, focuses on developing acoustic solutions for monitoring coral reefs. With coral reefs supporting about 30% of marine biodiversity and providing essential goods and services to nearly 1 billion people, Reef Pulse addresses the growing threat of reef degradation due to human activities. The company uses soundscape recording combined with signal processing and AI algorithms to continuously and automatically assess the ecological health of coral reef sites, offering a more effective alternative to traditional, subjective visual surveys.
During my internship, I contributed to the development of a deep learning model for bioacoustic analysis, aimed at automating the detection of key marine species and human-related noise pollution. I was involved in the annotation and validation of sound data collected from coral reef ecosystems in Mayotte and Guadeloupe. This process led to the successful detection of parrotfish and grouper species, and the development of a balanced and representative dataset for improved model performance. Additionally, the model was adapted to detect various sound classes, including boats, whales, and environmental noise, with further work underway on fine-tuning these detections.
What I particularly appreciated during my internship was the autonomy to explore my ideas and the opportunity to work on complex, meaningful tasks. I was involved in various aspects of the project, from AI and database management to backend development and code reviewing, which allowed me to broaden my skills.
————————————————————————————————-
Haolong Li
Description
Logitech is a leading global provider of product solutions that connect people with the digital world. Logitech provides hardware and software for various computers, communication, and entertainment platforms. During my internship at Logitech, I developed a desktop app using JavaScript, Python and C++ to track and analyze mouse in real time for personalized DPI recommendations, reducing user configuration time by 90%.
Besides the technically challenging project (and of course, a rewarding addition to my professional experience), Logitech also provides a young, energetic, and fun working environment. I had great colleagues, and we organized lots of interesting events. I also received professional and supporting supervision from my supervisors and learned things beyond engineering.
I definitely recommend Logitech if you are looking for a self-owned exploratory project.
————————————————————————————————-
Robin Faro
Description
Paradigma SpA is a software development company specialized in cloud-based web and mobile applications, user experience and interface design, and communication strategies. During my internship, I worked on developing an AI-powered tool to automate the review process of GitHub pull requests (PRs). My project involved several key phases, starting with the extraction of relevant data from PRs, including modified code, commit messages, and reviewer comments. Using this data, I generated a structured dataset by formulating targeted prompts and leveraging AI to produce high-quality reference reviews, ensuring consistency with the company’s standards.
Based on the Qwen architecture, a local model was trained on the previously generated dataset, leveraging techniques such as quantization, LoRA (Low-Rank Adaptation), and RoPE Scaling, which improved efficiency and reduced computational costs. Subsequently, the model was deployed on an AWS EC2 instance with GPU acceleration. I integrated the system into GitHub Actions, allowing automatic triggering of AI-generated reviews upon PR creation.
In the final phase, I developed a post-training pipeline to enhance the model’s performance on emerging programming languages and frameworks. This involved extracting documentation from web sources, generating synthetic training data, and fine-tuning the model while preserving its initial code-review capabilities.
This internship provided me with valuable skills in AI model training and deployment, cloud computing, and software development workflows. Additionally, working in an Agile environment allowed me to understand industry-standard methodologies and the practical challenges of integrating AI into real-world software engineering processes.
————————————————————————————————-
Massimo Rizzuto
Description
CHUV (Centre Hospitalier Universitaire Vaudois) is one of Switzerland’s leading university hospitals, renowned for its advanced research in biomedical data science and artificial intelligence applied to healthcare. The Biomedical Data Science Center (BDSC), in collaboration with the University of Lausanne, develops innovative solutions for privacy-preserving data science.
During my internship at CHUV, I explored synthetic data generation techniques with privacy guarantees, aiming to enhance data quality for research while preserving patient confidentiality. In the initial phase, I assessed the feasibility of various algorithms by testing them on public datasets to evaluate the trade-off between utility and privacy. At a later stage, I applied these techniques to real CHUV datasets, addressing challenges related to data preparation and quality. My main contributions involved implementing synthetic data generation models, analyzing their impact on accuracy and fairness metrics, and evaluating their potential applications in clinical decision support.
————————————————————————————————-
Paul Devianne
Description
General Electric HealthCare is an American health technology company operating in more than 160 countries in the world. The company specializes in all kinds of medical imaging devices / interventional equipment / monitoring devices. Their machines serve more than one billion patients every year with over 50 000 employees around the world. In Buc, France, my department is specialized in the development and production of Mammography systems. Breast cancer is the second deadliest cancer worldwide, with more than 670k deaths each year. Improving cancer detectability is a very promising path towards early diagnosis and reduced examination time.
My work there was to improve the Digital Breast Tomosynthesis (DBT) image quality using Deep Learning solutions. DBT is an advanced mammogram technique that projects X-rays from different angles to reconstruct a 3D volume of the breast tissues. In the last ten years, it has proven its efficiency in detecting breast cancer compared to classical 2D mammograms, both in terms of sensitivity and specificity. Recent advances in Deep Learning, with increasing hardware capabilities, has demonstrated great potential in improving Medical Image processing. By exploring the current limits of the DBT device and algorithms, we were able to identify areas of improvement and build a Deep Learning model capable of improving the image quality of the images. This progress is very promising to help clinicians perform more accurate and faster diagnosis.
————————————————————————————————-
Marianna Dell’Otto
Description
Corintis is a Swiss technology startup founded in 2022 and based in Lausanne, focused on solving heat management issues in semiconductor chips. With computing power demand rising due to advancements in AI, climate modeling, and drug discovery, chipmakers face limitations in improving performance due to excessive heat generation. Corintis has developed a revolutionary cooling solution that embeds microscopic channels inside chips, achieving significantly higher heat extraction rates and energy efficiency. This innovation reduces the energy consumption of cooling systems, which account for a significant portion of data center electricity use, and supports sustainable computing. The company fosters a dynamic, collaborative, and international work environment. During my internship in the Software Numerical team, I contributed to improving the software (based on Finite Element discretization for solving Partial Differential Equations) that simulates cooling processes. My tasks involved adding some features to the code to enhance simulation accuracy and exploring turbulence modeling for better heat transfer predictions. The internship strengthened my technical skills, particularly in collaborative software development, and offered valuable exposure to real-world problem-solving in an innovative startup.
Carlos Collado Capell
Description
The European Space Agency (ESA) is Europe’s gateway to space. Founded in 1975, it is a 23-
member intergovernmental body devoted to space exploration. My internship took place
at ESRIN, ESA’s Centre for Earth Observation, located in Frascati, Italy, near Rome. Within
ESRIN, I worked at Φ-lab’s Explore Office, a research unit focused on AI for Earth
Observation (EO), covering areas such as direct EO applications (weather modeling, wildfire
prediction, or sea ice monitoring), dataset creation, and Foundation Models, among others.
During my internship, I worked on developing a Foundation Model (FM) for onboard AI
processing on Φsat-2, a 6U CubeSat launched in August 2024. Φsat-2 is designed to
demonstrate real-time AI capabilities in space, especially for time sensitive applications.
Throughout the internship, I trained my own Foundation Model and evaluated the
performance of existing FMs on Φsat-2 data. My work involved overcoming the challenges
of onboard processing, including memory limitations and hardware-specific constraints. I
also collaborated with leading experts in Foundation Models for EO and onboard AI to refine
and optimize the approach. The project concluded with a pretrained backbone which will
now start to be benchmarked in different datasets relevant to onboard AI.
Younis Bouzar
Description
IMDM SA is an international consulting firm specializing in asset management for infrastructure and energy, with a strong focus on railway networks. The company leverages advanced simulation tools to audit and optimize railway and road networks, helping improve asset management processes. The work environment at IMDM SA is collaborative, fostering both independent problem-solving and teamwork.
During my internship at IMDM SA, I worked on creating and optimizing a predictive maintenance strategy for railway infrastructure, in collaboration with SBB Infrastructure. The project focused on developing a data-driven maintenance approach by analysing large datasets and designing an optimization model to enhance maintenance efficiency. The model incorporated various constraints and uncertainties to improve intervention planning and reduce costs. The results provided insights into optimizing maintenance schedules while considering operational constraints.
The internship was an enriching experience that allowed me to apply my knowledge to real-world industrial challenges. Throughout the internship, I improved my skills in mathematical modelling, data analysis, and project management, while also gaining experience in effectively communicating technical results in a professional environment.
Antonio Tirotta
turing of CNC (Computer Numerical Control) machines for high-precision industrial grinding
of cutting tools. The headquarters are located in Le Landeron in the Canton of Neuchˆatel, but
there is also an innovation cell on the EPFL campus at the Innovation Park, where I spent most
of my internship.
The company offers with the machines its own software package, which provides a set of
utilities for the design, numerical testing and visualization of the cutting tools produced by
their clients. Specifically the goal of the software is to compute grinding trajectories to produce
cutting tools with the desired geometry and manage their serial production.
The primary goal of the project was to study and design numerical methods for computing
trajectories that minimize the jerk, i.e. the third derivative of the position since, in numerically
controlled machining, an high-quality surface finish is very important. However, due to the
complexity of the problem, I worked on the first derivative. Specifically, I have developed a
2D model where to conduct the investigation and, after having added an optimization routine,
I analyzed the different landscapes of optimization both experimentally and theoretically de-
pending on some free parameters. At the end, I implemented a C# prototype of the 2D model
where all the findings of the above study were incorporated.
During my internship, I had the chance to work in a stimulating and dynamic environment
that allowed me to improve both technical and soft skills. I would like to express my gratitude
to my supervisors and all the colleagues that I met during this experience.
Melchior Carrier
Description
Afflux is a Swiss consulting firm established in 2022, based at the Innovation Park of the Ecole Polytechnique Fédérale de Lausanne (EPFL). The company specializes in leveraging artificial intelligence and digital twin technology to optimize processes in industrial and medical sectors. The main goal of afflux is to enable their clients to make confident decisions based on simulation supported by real data. Its core activities include:
- 3D Digital Twin Creation: Using FlexSim software to model and simulate processes.
- Custom Simulation/Optimization Development: Designing tailored simulations to identify optimal solutions using custom AI algorithms.
The 3D models enable the simulation of future or existing processes, allowing for the testing of different hypotheses, the anticipation of potential problems, and the identification of critical factors limiting system performance. The custom AI tools, combined with tailored simulations, facilitate the exploration of thousands of solutions to find the optimal one. My internship involved contributing to two ongoing projects of tailored Simulation and Optimization in industry. This required utilizing mathematical and algorithmic tools to develop an AI algorithm and various KPIs.
My two main projects were:
- DSR project: Created a Tableau dashboard to analyze the performance of a previous afflux project, combined with statistical analysis.
- LB project: Implemented AI algorithms and collaborated on the creation of a simulation for a shop floor.
Theses two projects were done using C++.
Maxence Hofer
Desciption
—————————————————————————————————————
Francesca Bettinelli
Description
Hitachi Energy is a technology company that provides high-voltage equipment, transform-
ers, and services for renewable energy. Headquartered in Zurich, the company employs
45,000 people across 60 countries and produces a business volume of 13 billion USD.
During my internship at the Baden-D¨attwil research center, I was part of the Systems
and Automation team, whose mission is to develop digital solutions in power systems,
software design, and artificial intelligence. My project aimed to assess the applicability
of large foundation models for weather – such as FourCastNet or Pangu-Weather – to
renewable energy forecasting.
More specifically, in the first part of the project, I selected some open-source, pre-trained
foundation models for weather from the literature. Then, I trained some deep learning
models to predict the wind and photovoltaic energy production based on the weather
state in a region of interest. Finally, I used AI-generated weather forecasts (an example
is shown in Figure 1) as input to these models for a specific use case involving a hybrid
wind-solar power plant.
My work resulted in an internal technical report, a presentation to the team, and a paper
for future publication. This internship was a valuable learning experience, allowing me to
improve my technical and programming skills while taking full ownership of my project
from the very beginning.
Figure 1: Eastward wind component at 10 m altitude over a latitude-longitude grid
covering the Earth’s surface.
—————————————————————————————————————
Anya-Aurore Mauron
Description
My internship was hosted by IDUN Technologies, a pioneer in the development of in-ear EEG earbuds dedicated to the monitoring and interpretation of brain waves. Established in 2017 as a spin-off from the Federal Institute of Technology (ETH) in Zürich, IDUN Technologies is committed to enhancing individuals’ self-awareness by offering real-time, objective measurements of brain activity.
Legend: Raw VS Filtered data. Labels correspond to the absolute angles of the gaze of the subject. Calibration time is necessary for the filter to be efficient.
—————————————————————————————————————
Francesco Fainello
Description
Logitech is a leading company in computer peripherals and software, known for its innovative products that enhance user experiences across different platforms. As the top employer of EPFL students, Logitech provides opportunities for interns to work on cutting-edge projects. This particular internship, supervised by the Audio ML team and the CTO office, focused on exploring deep learning solutions for speech separation in video conferencing products. The aim was to develop a deep learning framework capable of enhancing the user experience by effectively separating overlapping speech and improving audio quality in real-time. This involved developing a deep neural network to process multi-channel audio signals, extracting both spectral and spatial information to distinguish between different speakers. The work culminated in the creation of a Python-based model capable of separating and enhancing speech, even in scenarios with an unknown number of speakers. Through continuous support and weekly meetings, I was not only able to deepen technical skills in programming and deep learning but also to refine my communication skills through regular presentations and interactions within a structured corporate environment, providing valuable support and insights into organizational dynamics.
—————————————————————————————————————
Giacomo Mossinelli
Description
Procter&Gamble is the world’s largest consumer goods company, made of 100 000 employees with a turnover of $82.0 billion. I did my internship in the R&D department of their German Innovation Center (GIC) in Schwalbach am Taunus, near Frankfurt am Main.
During my 6 months long internship project I had the opportunity to focus on modeling and simulation tools to assist and improve the development of products made of Airlaid materials (figure 1).
More specifically, in my project I focused on the improvement and optimization of the numerical model for one of the most important metrics to describe the final product performance, so that it could be implemented together with the other already returned metrics. The result thus obtained is to have a single simulation that provides the fundamental information about the product without the need to invest resources in tests and experiments. The final step was to use what was built to create a virtual twin of a very intricate and relevant test method, which is composed of many different measurements. The outputs of this virtual experiment can finally be used to take developmental and business decisions.
Right from the beginning of the internship, I was charged with many responsibilities and project leadership. This allowed me to learn a lot from a technical point of view but, more importantly, from a professional, relational and organizational point of view. Indeed, to bring the project to completion it has been necessary to interact and collaborate with colleagues with different roles and backgrounds, thus making the social and communication component as crucial as the technical and engineering one.
Moreover, the opportunity to conduct experiments in the laboratory following the directions and tips of experts, with the aim of validating the predictions of numerical simulations, allowed for a practical approach in addition to the more theoretical and modeling approach of the strictly modeling part.
Figure 1: Portion of Airlaids manufacturing process
————————————————————————————————————–
Kseniia Shevchenko
Description
Alpiq is one of the main electricity producers in Switzerland with a large majority of its production coming from hydropower plants. Alpiq also has varied portfolio of nuclear, gas-fired, wind, solar power plants located throughout Europe. During my internship I worked in the Innovation and Projects team whose goal is to bring innovative digital solutions into production to optimize the operation of hydropower plants.
During my internship I worked on Alpiq’s predictive maintenance project which is aimed at preventing failures on the power plant detecting early detection of precursors. The core of the project is machine learning models trained for anomaly detection. I worked on multiple tasks, such as hypothesis testing, development and integration of new features (metrics) and physics-based models. The developed metric was based on statistical indicators derived from the residuals of the prediction model and proved to enriches and complements the existing functionality of the project. The goals of the physics-based model was to directly model the thermal behavior of and element of the machine using a discretized dynamical system.
During the internship I enhanced my skills in machine learning and software engineering. The internship allowed me to focus on specific domain problem – in maintenance of hydropower equipment and see how to integrate prior physics knowledge into machine learning models as well as how to interpret the results according to the construction and functioning of power plant.
This internship provided me an excellent learning experience and was very valuable both in improving and getting new technical skills and collaboration and communication in professional context in industry.
————————————————————————————————–
Sam Jegou
Description
APCO Technologies specializes in designing and manufacturing high quality mechanical and electromechanical equipment for space, energy and industry applications. With around 400 employees, APCO focuses on providing innovative solutions and concepts, from the engineering phase to assembly and final testing, as well as on-site assistance.
During this internship I developed an algorithm that recommends collision avoidance maneuvers to satellites. These maneuvers are optimized to use as less propellant as possible, but still reaching low-risk criteria (distance between the two satellites and probability of collision). Another type of maneuver proposed is to change the ballistic coefficient of the satellite (via the solar panels orientation for instance) to modify its trajectory, as shown in the figure.
