Partner Challenges

Full Description

OptiKonf's platform uses Symbolic AI to process complex specifications precisely, even when given in non-standard input formats. It can process imperative code (e.g., a sub-set of Python syntax) and transform it into configurators — crucial for processing existing configurators found in many companies. While this makes symbolic AI accessible to programming experts, OptiKonf strives to build an even more accessible platform.

The company is researching agentic programming workflows (e.g., external APIs with user-supplied keys, or budgets bought through the platform) that still exploit the symbolic back-end. Generated Python source can always be processed to extract selectable configurator "knobs", optimisation targets, dead code, etc. This information should be exposed to the agentic coding loop, so users are never "vibing into the void" but stay connected to the domain problem they are modelling — able to ask the agent questions while it calls the platform in the background to compile code and execute configurators.

Participants receive API access to the platform and personalised guidance throughout. The team supports any problems that arise and values feedback. Deployment is handled by OptiKonf — the challenge focuses on collectively gathering experience and seeing how far the group can push the workflow.

Data & Resources

• API keys for the OptiKonf platform provided
• No confidentiality required (isolated tenant)
• Outputs may be shared publicly

Company

Kleinstunternehmen · 2 employees, 2 founders · Linz, Austria

optikonf.com ↗

Full Description

SAP builds software that helps companies run well. The Signavio business unit provides a process intelligence product suite as part of SAP's business transformation capabilities. Signavio operates at the level where businesses are actually run — the process level — offering tools and knowledge to get insights into how processes run and adapt them to drive change. Over 50 years, SAP has accumulated knowledge of potential process improvements, but whether a specific improvement will work in a given organisational reality requires decision support beyond this knowledge alone.

Business process simulation provides that decision support by learning a simulation model from historical execution data and modifying it to reflect hypothetical changes — enabling what-if analyses before committing to redesigns, automation, or resource reallocation. In practice, however, these approaches still do not produce simulations accurate enough to rely on.

A core reason is that the data used to discover simulations does not capture everything driving process execution: processes are variable because humans execute them, and they are shaped by external context (workload, customer behaviour, supplier delays, organisational dynamics) that is only partially recorded in the ERP. Promising directions include integrating context from outside the ERP, modelling resource behaviour under varying conditions, and quantifying how trustworthy a simulation result is — through new paradigms or hybridisations.

The relevance grows as agents take a more central role: SAP's active research on agentic process management explores autonomous agents, and sufficiently accurate simulation is key to understanding their behaviour and consequences in advance.

Data & Resources

Openly available process logs at tf-pm.org/resources/logs ↗ paired with the state-of-the-art references below.

• Kirchdorfer et al. (2025). Discovering multi-agent systems for resource-centric business process simulation. Process Science, 2(1), 4.
• Chapela-Campa et al. (2025). SIMOD: automated discovery of business process simulation models. SoftwareX, 30, 102157.
• Camargo et al. (2023). Learning business process simulation models: a hybrid process mining and deep learning approach. Information Systems, 117, 102248.

Company

sap.com ↗

Full Description

AI systems are evolving from isolated models into interacting ecosystems of autonomous agents, humans, digital infrastructure, and physical environments. Risk is no longer located solely inside a single model — it emerges from the behaviour of the system as a whole, including how software agents, physical systems, and people interact over time.

Consider an autonomous construction site where multiple machines collaborate with human operators to move material. Every decision changes the environment itself: terrain is reshaped, materials move, traffic patterns evolve, and new decisions must adapt to the updated world. New safety rules, optimisation objectives, or operating policies influence how autonomous agents coordinate. Small changes to either the physical models or the governing rules may produce system-wide effects that are difficult to anticipate.

Building trustworthy autonomous systems requires understanding both the physical processes and the interactions between multiple adaptive agents. Physics-based digital twins provide realistic virtual environments for training and validating autonomous machines, while observability and governance methods help us understand how changes to rules, incentives, or control strategies affect overall system behaviour. Together, these perspectives aim to provide confidence that autonomous systems remain safe and stable as they continuously adapt.

The challenge sits at the intersection of Physical AI, physics-based simulation, multi-agent systems, and AI governance, and is presented jointly by Algoryx and cleAIr under the WASP umbrella.

Domains

Physical AI · Physics-based simulation · Multi-agent systems · AI observability & governance · Autonomous machines · Construction · Space systems

Data & Resources

• Data: No
• Examples: Yes
• Supporting materials: No
• No confidentiality or IP constraints
• Outputs may be shared publicly

Companies

Algoryx — Physical AI & Simulation · algoryx.se ↗

cleAIr — IT Research · cleair.ai ↗

Full Description

LLMs generate SPARQL queries against knowledge graphs but struggle with irrelevant schema noise. Ontology subsetting — including only the classes and properties actually instantiated in the knowledge graph — reduces the token load and can improve the quality of results.

Complex ontologies like CIM (the electrical grid standard) contain more than 900 classes and 5,500 properties, creating excessive token overhead for LLMs. Organisations need methods to make ontologies "LLM-friendly" without sacrificing precision, so that agentic AI solutions can be deployed within days rather than quarters.

Key objectives: reduce massive schemas to actively-used terms only; improve LLM accuracy in generating correct queries; and remove redundant definitions and unused enumerations.

Data & Resources

• Dataset: STATNETT (Norwegian grid operator) — 100+ named graphs, largest containing ~2 billion statements
• Stack: GraphDB, SPARQL, OWL RL reasoning, GeoSPARQL, timeseries integration
• Live demo: CIMOn Grid Semantic Digital Twin chatbot ↗

Company

graphwise.ai ↗

Full Description

Visual query answering and netlist generation are meant to automate the transition from visual circuit designs to digital netlists in enterprise systems. Current approaches — standard object detection, OCR, and general-purpose Multimodal LLMs — still fall short of the topological precision needed in practice.

The challenge is to close the topological accuracy gap, potentially by drawing from neurosymbolic AI, graph neural networks, or knowledge-guided learning. This becomes increasingly important as automated agents take on greater roles in hardware engineering workflows, where errors in netlist extraction can propagate into downstream design decisions.

Domains

Electronics Design · Industrial AI · Visual Reasoning · Multimodal LLMs · Neurosymbolic AI

Company

Bosch Center for Artificial Intelligence (BCAI)

bosch-ai.com ↗