Challenges 2026 - MAKEathon

Challenge 1. Token Consumption of RAG systems In a recent Graphwise internal benchmark experiment, GraphRAG was compared with alternative retrieval methods for answering competency questions over large corpora. The baseline test architecture utilized an extensive knowledge graph integrated with two primary datasets: 1) US public company records—including SEC EDGAR filings, annual reports, proxy statements, patent filings, and ticker registries for entities like Microsoft, Nike, Apple, Oracle, Johnson & Johnson, Alphabet, and Pfizer; and 2) a health plan dataset encompassing policy rules such as copays, coinsurance, preventive coverage, mental health care, IVF, and durable medical equipment. Early evaluations yielded encouraging outcomes regarding token usage efficiency, along with improvements in precision, thoroughness, and auditability. To carry out the challenge, we outline a baseline benchmarking framework to evaluate GraphRAG relative to direct prompt document injection and standard vector-based RAG. The evaluation will draw from either US public company disclosures (a) or a public health dataset (b), testing each approach against a structured suite of 20 competency questions across three distinct complexity tiers: single-fact, nested-fact, and multi-hop queries. The experimental setup leverages pre-existing datasets, models hosted within Graphwise GraphDB, and competency questions. To support various testing configurations—from standard Vector RAG and targeted prompt processing to full-scale GraphRAG—we deploy a dedicated Graphwise GraphRAG instance. The evaluation workflow integrates custom procedural hooks and script extensions directly into GraphRAG, operating alongside Graphwise GraphRAG’s native assessment framework to execute the benchmarking process. Outcomes: Primary Objective: Benchmark token consumption across the various configuration settings. If time permits, evaluate key metrics such as accuracy, completeness, and traceability. Utilize the recently launched Graphwise GraphRAG Evaluation Framework.
Challenge 2. Event-Driven Agents for Financing Decisions Many AI agents today are launched through chat interfaces. This use case aims to develop an event-driven agentic AI system that is automatically triggered by external events such as emails, market information, system notifications, or workflow updates. The agent analyzes the event, derives relevant information, and autonomously takes actions, generates recommendations, or initiates follow-up processes. One possible scenario is supporting investment decisions based on market or portfolio events. Potential Data Sources: Emails Event and system logs Workflow data Market and business information Expected Outcome: A prototype demonstrating how agentic AI can proactively support business processes through event-driven automation and decision support.
Challenge 3. AI-Powered Regulatory Monitoring for an Energy Utility Energy utilities need to continuously monitor legal and regulatory developments across various public sources. Today, identifying relevant changes requires considerable manual research. The challenge is to develop an AI-powered prototype that monitors selected Swiss legal and regulatory sources, detects new or changed information, and assesses its potential relevance for an energy utility. Possible workflow: Public sources → Change detection → AI-based relevance assessment → Structured result / notification. The prototype should help answer: What has changed? Is it relevant to an energy utility and why? Which topic or business area could be affected? Are there important deadlines? What is the original source? Possible sources include ElCom, federal and cantonal authorities, consultations and official publications. No confidential company or personal data is required. The goal is a proof of concept demonstrating how AI and domain knowledge can reduce manual monitoring effort while keeping results transparent and traceable to their original sources.
Challenge 4. Push Switzerland’s Truly Open LLM Further Apertus is Switzerland’s truly open large language model (LLM), developed by ETH Zurich and EPFL within the Swiss AI Initiative and trained on CSCS infrastructure. The goal is to adapt, specialize, re-structure, or improve Apertus and demonstrate a measurable business impact. Five Challenge Paths Choose one of the five paths below, or combine several. Fine-tuning, distillation, architectural changes, RLVR, quantization, sparse routing, new inference strategies, and agentic harnesses are all valid approaches. Path 1 – Bring Apertus into Business Build an agentic harness around Apertus for realistic office, knowledge-work, or coding tasks in a Swiss or multilingual business context. Teams may improve the harness, tool schemas, post-training, or the model itself but should compare the result against a clear Apertus baseline. Resources: BFCL; OfficeBench; WorkArena; SWE-bench Verified; Evaluation: end-to-end task success, tool-call correctness, recovery from failures, latency, and resource use. Path 2 – Turn Apertus into an Embedding Model Transform Apertus from a generative decoder into a retrieval or sentence-embedding model. Explore bidirectional attention, pooling, contrastive learning, Matryoshka representations, multilingual alignment, or task-specific post-training. An existing Apertus adaptation provides a concrete starting point and demonstrates one route to bidirectional retrieval. Swiss multilingual data such as VotingBooklets and SwissGov-RSD can support training or alignment, while VotingBooklets-Diamond can be held out for evaluation. Resources: Apertus Swiss Embed; VotingBooklets-v1; VotingBooklets-Diamond-v1; SwissGov-RSD; MTEB; Evaluation: Recall@k, nDCG, MRR, embedding size, inference speed, and multilingual retrieval quality. Path 3 – Build a Jev-like Structured Decision Model from Apertus Replicate Jev-like structured decision functionality using Apertus, either by modifying Apertus itself or distilling it into a smaller Apertus-derived model. Instead of generating prose, the model should consume text or application state together with runtime-defined typed questions and permitted answers, then return software-consumable decisions with probabilities and confidence information. GLiNER2.5- Decide provides an open-weight reference for schema-defined classification, routing, and triage. Possible approaches include a bidirectional or encoder-style Apertus conversion, decision or scoring heads, contrastive or classification post-training, teacher-student distillation, and constrained joint decoding for related decisions. Resources: Jev; GLiNER2.5-Decide; fast-decisions; CLINC-OOS; Banking77; Evaluation: exact-match decision quality across changing schemas, calibration (for example Brier score or ECE), latency, throughput, memory, and CPU/GPU viability. Path 4 – Improve Reasoning with Ontology-Verified RLVR Improve Apertus reasoning with reinforcement learning using verifiable rewards (RLVR) derived from knowledge graphs and ontologies. Construct tasks for which correctness can be checked automatically by executing SPARQL, validating graph paths or relations, checking ontology constraints, or comparing against deterministic program outputs. Wikidata, DBpedia, KQA Pro, or a domain ontology can provide structured knowledge and supervision. A practical route is to generate reasoning problems from the graph, establish a baseline, optionally use supervised fine-tuning for warm-up, and then apply RLVR with an explicit verifier. Resources: RLVR-Ontology example; Wikidata dumps; DBpedia resources; KQA Pro; Evaluation: answer accuracy, verifier pass rate, graph or constraint validity, robustness to unseen compositions, and changes relative to the base model. Path 5 – Build a Domain Expert Model That Knows Its Boundaries Specialize Apertus for a clearly defined domain while explicitly teaching it when not to answer. Choose a domain with suitable open data, such as Swiss law, public administration, public health, science, or another well-scoped field. Improve in-domain performance through supervised or other post-training, retrieval-aware adaptation, or targeted data curation, and combine this with out-of-domain detection, calibrated confidence, refusal, or routing behavior. Build separate in-domain and out-of-domain evaluation sets so that specialization does not simply make the model more willing to answer everything. A strong result improves domain quality while making boundaries more reliable. Resources: Swiss Leading Decisions; Swiss Open Government Data; FOPH/BAG Infectious Diseases Dashboard; Evaluation: in-domain quality and faithfulness, out-of-domain rejection or routing, calibration, false refusal rate, and robustness near the domain boundary. Starting Points, Data, and Infrastructure Teams may use official Apertus releases or experimental derivatives. The following starting points are particularly useful and are listed with explicit URLs: Apertus 1.5 8B text-only text-only derivative packaged as a standard ApertusForCausalLM checkpoint for fine-tuning, distillation, inference, and architecture experiments. Apertus Swiss Embed; bidirectional retrieval adaptation using Swiss multilingual training and evaluation data. Swiss AI Initiative models; official Apertus models, collections, and related resources. FHNW access: Swiss AI Research Platform Apertus inference can be provided to participants using an @fhnw.ch account; this is useful for prototyping and evaluation without hosting the full model locally. CSCS compute FHNW students using an @students.fhnw.ch address can use CSCS compute allocated to FHNW, subject to onboarding and quota. Contact the challenge owner early if GPU compute is required. Selected Data and Benchmark Links These concrete sources are intended as starting points. Teams may use other suitable open datasets and benchmarks, but should always check licenses, terms of use, data protection, and whether the source fits the intended task. BFCL V4 tool/function calling and agentic evaluation; relevant to Path 1. OfficeBench multi-application office automation workflows; relevant to Path 1. VotingBooklets-v1 large Swiss parallel corpus in German, French, Italian, and Romansh; relevant to Path 2. VotingBooklets-Diamond-v1 manually corrected evaluation subset; useful as held-out multilingual evaluation for Path 2. SwissGov-RSD human-annotated multilingual Swiss government documents; relevant to Path 2. fastino/fast-decisions 5,100 held-out examples across 17 operational decision domains; relevant to Path 3. GLiNER2.5-Decide 340M open-weight reference model for schema-defined decisions; relevant to Path 3. CLINC-OOS intent classification with explicit out-of-scope examples; relevant to Paths 3 and 6. Banking77 13,083 banking queries across 77 fine-grained intents; relevant to Path 3. KQA Pro complex knowledge-graph QA with executable SPARQL/program supervision; relevant to Path 4. RLVR-Ontology example worked example of an ontology-oriented RLVR setup; relevant to Path 4. Swiss Leading Decisions 21k multilingual Swiss Federal Supreme Court cases; relevant to Path 6. opendata.swiss Swiss Open Government Data catalog for domainspecific data; relevant to Path 6. FOPH/BAG Infectious Diseases Dashboard Swiss public-health surveillance data and dashboards; relevant to Path 6. Starting Points, Data, and Infrastructure A working Apertus-based prototype or model adaptation in a reproducible repository. A clear baseline, technical hypothesis, and concise description of the approach. At least one quantitative or executable evaluation against the baseline. Results, limitations, and the main quality, efficiency, or reliability trade-offs. Evaluation will emphasize technical originality, evidence of improvement, reproducibility, practical relevance, and responsible use of data and compute. The strongest project is not necessarily the one using the largest model or the most compute.
Challenge 5. AI Adoption Advisor for SMEs Small and medium-sized enterprises (SMEs) form the backbone of the Swiss economy, yet many of them struggle to move from curiosity about AI to real, value-generating adoption. Where should a company start? Is it ready for a first use case, or does it first need to get its data, skills and infrastructure in order? And how can management be convinced that the next investment is worth it? Large corporations can hire consultants to answer these questions; most SMEs cannot. In the applied research project AI4SMEs, FHNW has developed methods that help SMEs assess where they stand on their AI adoption journey and plan their next steps in a structured way. With this challenge, we invite participants to turn this knowledge into an intelligent advisor that any SME can use. The system should engage with a company, for example through a guided conversation, a questionnaire or the analysis of documents the company provides, and determine its current AI maturity stage across relevant dimensions such as strategy, data, processes, technology, skills and governance. Based on this assessment, the advisor should recommend concrete next steps that fit the company’s stage, size and industry. Crucially, recommendations must be grounded in the AI4SMEs methods rather than generic LLM knowledge, and they must be justified from a business perspective: each suggestion should come with an estimate of the expected return on investment, the KPIs the company can use to track success, and the effort, risks and prerequisites involved. For example, a 40-person manufacturing company might describe how it handles customer orders and quality control today. The advisor could classify the company’s maturity level and explain why, identify gaps that block progress, such as scattered production data or missing internal AI skills, propose a prioritized roadmap, for instance a document-processing pilot for incoming orders before tackling predictive quality, and back each step with a business case showing expected cost savings, payback period and KPIs like order processing time or defect rate. Participants are encouraged to explore how the AI4SMEs methods can be formally represented, for instance as a knowledge graph or ontology, and combined with LLM reasoning so that the advisor’s assessments and recommendations remain explainable, traceable and consistent. Multi-agent architectures, where specialized agents handle assessment, recommendation and business case calculation, are equally welcome. Materials describing the AI4SMEs methods will be made available to participating teams.
Challenge 6. DentalGuard – Explainable AI for Early Detection of Oral Discomfort in Cats Cats are very good at hiding pain, meaning dental and oral problems can remain unnoticed until they become more severe. AI-Tails is developing a smart feeding and drinking station that pas-sively observes cats during their everyday routines. The challenge is to explore whether these interactions can be used to identify early signs of possible oral discomfort. Participants are invited to develop an explainable AI prototype that analyses feeding behavior such as difficulty grasping or dropping food, chewing patterns, head movements, meal dynamics and pain-related facial cues. Instead of comparing a cat only with population averages, the sys-tem should learn the individual cat’s normal behavior and identify meaningful changes over time. A particular challenge is incomplete observation: outdoor cats may eat or spend significant time away from the station. The prototype should therefore estimate how much reliable information is available, adapt its confidence accordingly and recognize when there is insufficient evidence to make a reliable assessment. The goal is to create an AI solution that can detect early-warning signals and process them to deduce and explain oral discomfort in cats. Such oral Challenge question: How can we identify and explain oral discomfort in cats from everyday feeding behavior – even when we observe only part of the cat’s day?
Challenge 7. Build an AI-powered Operating Model with IBM watsonx Orchestrate Organizations are increasingly challenged by the growing number of communication channels, systems, and tasks that employees need to manage every day. Important actions are often hidden across emails, messages, social networks, collaboration tools, and business applications. IBM watsonx Orchestrate enables organizations to build AI-powered assistants that can connect to different information and communication sources, identify relevant actions, and help users manage their daily work more effectively. Your challenge is to design and prototype an AI-powered operating model that acts as a digital work assistant for employees. The solution should be able to collect information from multiple communication channels such as email, LinkedIn, collaboration tools, and other business systems, identify relevant actions, automatically generate tasks, and help users manage their work. The AI assistant should support employees in tracking and managing tasks through different stages such as “Not Started”, “Started”, and “Done”, while providing a simple and intuitive user experience. Teams are encouraged to use their creativity within these boundaries to design an innovative solution that demonstrates how AI can help employees stay organized, prioritize work, and increase productivity across multiple communication streams. Proof of concept: a working prototype, a short demonstration of the user journey, and a description of the business value delivered by the solution. Outstanding solutions may have the opportunity to contribute to a real-world implementation after the MAKEathon. Depending on the outcome and feasibility, the resulting solution could also be published as an out-of-the-box asset within the IBM Cloud SaaS environment. Students involved in outstanding solutions may additionally have the opportunity to continue contributing to the development of the solution after the MAKEathon.

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.

X/Twitter
This field is for validation purposes and should be left unchanged.
Name*
First Last
Email*
Website
Challenge*

Looking for inspiration or curious about what to expect?

Current challenges haven’t been finalized yet, don’t worry — you can still explore previous Makeathon editions to get a feel for the types of problems we’ve tackled in the past.
Whether you’re here to submit a challenge or just browsing, take a look at some of the exciting ideas from past years below.