| Challenge 1. Explainable-AI by PostFinance As AI methods become more established in business processes, both regulators and business users are becoming increasingly interested in model explainability. There are a few standard technical metrics, such as Shapley values, that are often used in this context, and which we already use in several product groups, such as customer personalization, fraud prevention, and money laundering detection. However, these metrics can still be difficult to interpret quickly in a business setting far removed from the technical model implementation. We postulate that an LLM-assisted product, which combines the Shapley values corresponding to a given model score, the technical documentation of a model including feature and target definitions, and the power of a modern generative AI model could provide relevant explainability interpretations at the level of detail required by a given application. We will provide a set of example data corresponding to a specific business application. Your task will be to build a model, extract relevant explainability metrics, and build a product that summarizes those metrics in German or English language for the business user. | PostFinance |
| Challenge 2. AI Assistant for Research Reproducibility Scientific papers often introduce or reference complex software, datasets, or experimental procedures that are crucial for reproducibility, but finding, understanding, and running these resources can be time-consuming and technically challenging. With this challenge, we invite to design and develop an AI-powered assistant that aids researchers in navigating this complexity. Using scientific knowledge graphs like SemOpenAlex, the AI assistant should help users discover relevant software, datasets, and resources from academic papers. The system should not only locate these resources but also assist in setting them up, explaining how they work, and — ideally — automating the process of running experiments. The complexity of the assistant could range from an exploratory tool that maps relationships between resources in a paper (e.g., software and datasets) to an advanced agent that uses reasoning and automation to fully set up and reproduce experiments. For example, given a software-related research topic, the assistant could retrieve relevant papers from SemOpenAlex and provide information and setup instructions for the artifacts produced in these papers. External links to the artifacts (e.g., on GitHub or Zenodo) can be obtained from analyzing metadata in SemOpenAlex and connected knowledge graphs like Linked Papers with Code or Wikidata. Based on the nature of the artifacts, the level of guidance from the (LLM-based) assistant may vary. For example, for a dataset the assistant could provide a summary, an overview of connected datasets, and produce a data sample. For a software library, the assistant could provide a description of its capabilities and how to apply it in the context of the research topic. Ideally, the assistant will provide additional guidance on how to apply these artifacts in an integrated setup. | Metaphacts |
| Challenge 3. Drone aircraft inspection Swiss Airlines is testing using drones to inspect the aircraft. Drones can be fast and could potentially inspect aircrafts at the airports or the hangars in matter of minutes from different angles (from above or under). But of course, speed is not the only requirement. We would need to strike an optimum balance between costs, speed, accuracy, and safety. Regarding safety, drones shall avoid collisions (aircraft and airport facilities) and people (airport staff and crews) and preventing technician’s accidental falls from heights. Link to video: https://www.youtube.com/watch?v=jbaDtxSBGHU Data: pictures taken and labelled by participants themselves at the MAKEathon on a mock-up airplane. We will provide one or several mock-up aircrafts. You will take some pictures with your smartphone (different angles, light conditions, different mobile cameras), draw some erasable markings on the plain and take more pictures. It requires creating a pipeline to 1) capture pictures/videos (with smartphone cameras) –> 2) segmentation to locate the aircraft (against the background or people) –> 3) (optional) positioning and collision avoidance navigation of the drone –> 4) classification to identify which part of the aircraft it is (tail, wings, etc.) –> 5) classification to decide healthy or damaged / anomaly –> 6) classification to a type of damage (lightning strikes, hail blows, etc. ) –> 7) quantify the damage (size, amount, angles, etc.) –> 8) logistic regression to a severity –> 9) decision on a recommendation (fly or repair) –> 10) generate a human-readable report –> 11) Q&A chatbot responding to questions on the report. | SWISS Airlines |
| Challenge 4. Multichannel customer support “Swiss all the way” is both our new guiding principle and a promise. SWISS offers Swiss service quality from A to Z and sees itself not only as an airline, but also as a host that provides comfortable travel experiences – with attention to detail, warmth and authenticity. We want to assist our customers on the phone while they are simultaneously booking, rebooking or solving a problem on the web site. We envision a voice assistant giving instructions on what to do and gets real-time feedback of user’s action on the web site. PoC: develop a proof of concept on a very simple web app will do. It shall be a generic solution applicable on other web sites. The solution shall combine language, speech-to-text and text-to-speech, etc. Data: description of different typical customer interactions (booking, flight information, irregularities, rebooking, luggage claims, etc.). Publicly available LLMs for language understanding. | SWISS Airlines |
| Challenge 5. Natural Conversation Existing voicebots make conversations with intelligent systems easy and hands-free. However, having natural conversations with them is still far from reality. In particular, one cannot expect graceful reactions to interrupting a bot. It would be nice however, if a bot could understand why a human is interrupting: if, for instance, the human is already aware of certain information, the bot may understand an interruption as a sign of such awareness and continue the conversation on a “higher level”. The challenge aims at solving this problem in general and/or for a more particular context. That particular context is given by educational games that students can use to simulate interviews with business stakeholders. When prompted with a question, the bots – in this case simulated stakeholders – will reveal certain predefined information in written form. However, we want to give the games a more lively touch by introducing speech and natural reactions of stakeholders as described above. There are two games (with corresponding dialog scripts) on which the approaches can be tested / developed. | IIS Research Group |
| Challenge 6. SpeedyBots Educational chatbots are increasingly being adopted because they are effective learning companions available 24/7. German-speaking train drivers from the railway company SOB (Schweizerische Südostbahn AG) would benefit greatly from an educational chatbot for their operations in Ticino, where they have to be able to communicate in a technical Italian language. However, creating and maintaining such an educational chatbot is challenging because it requires integrating domain knowledge and “Conversational AI” knowledge. A low-code application could address this challenge by allowing domain experts (i.e. language teachers in this case) to quickly create and maintain chatbots (hence SpeedyBot) without (or the minimum) intervention of “Conversational AI” experts. In this challenge, you are asked to conceive such a low-code application focusing on how to use AI technologies effectively, e.g., is LLM RAG with Knowledge Graph a suitable approach? What’s your suggestion? During the MAKEathon, the teaching material will be made available, and a domain expert will be at your disposal. | SOB |
| Challenge 7. Airport operator and the reinforcement agent In collaboration with the internationally present airport ground handling service provider Swissport, our research team has begun work on the development of AI and ML systems which will monitor operations, simulate future events and provide recommendations for optimization and disturbance mitigation. Aviation being as volatile as it is, Swissport’s operations are subject to a multitude of external variables which can have substantial negative impacts, leading to delays and customer dissatisfaction. These uncontrollable events are equally difficult to predict for human dispatchers, making choosing an appropriate course of action equally challenging. The envisioned solution reduces the mental load for human dispatchers by serving as an easily accessible monitoring system, providing not just real-time metrics but also prognoses and trajectories for different time-horizons. An AI system trained through observation of human dispatchers provides action recommendations, which can be accepted, rejected or adjusted by the human operator, creating an online learning environment for the agent while simultaneously supporting human decision-making. A high degree of human control is essential to the project, with the end goal of fostering trust and an environment of mutual learning between AI agents and human operators. Proof of concept: Develop a human-machine interface (or just a concept) to ensure communication between the airport operator and the reinforcement agent. Use a RAG approach or a self-developed approach. | SwissPort |
| Challenge 8. LLM for underwriters Underwriters at Baloise Insurance evaluate and analyze risk to determine the terms of coverage and pricing. This process involves reviewing large amounts of data, client applications, historical claims, market conditions, and industry trends, which can be time-consuming and prone to human error. The goal is to leverage a large language model (LLM) to help underwriters streamline their workflow, reduce manual effort, and improve accuracy. Proof of concept: Write an concept and a small RAG (LLM) for an automatic process that extract key insights from client applications, historical claims, policy documents, and external data sources like industry news or market reports. | Baloise |
Are you a company or a research institute with an AI challenge? Submit your challenge using the form below!
Submit Your Challenge
We will review your challenge and get back to you to inform you about the selection, rejection or to request further information.
