| Agentic AI | Agentic AI refers to autonomous or semi-autonomous software systems that use advanced AI techniques to perceive, make decisions, take actions, and achieve goals in their digital environments, often adapting to changing conditions. | |
| Generative AI in the Software Lifecycle | Generative AI may not only be useful for coders, but it can expand to all other actors in the software development lifecycle. Business logic, UI & UX, data layers, … important steps in the generation of all these aspects can be made by business experts starting from a mere description. | |
| Guardrails & Evaluations for RAG | Guardrails are techniques and tools ensuring AI output aligns with policies, preventing harmful content or data leaks. Evaluations, especially automated ones, assess the quality of RAG-based applications, enabling scalable, cost-efficient monitoring of metrics like relevance and accuracy. | |
| AI on sensitive data | AI on sensitive data involves applying AI techniques to confidential or personal data like medical or financial records. It can unlock valuable insights while necessitating strong privacy measures to protect individuals and comply with regulations. | |
| GraphRAG | GraphRAG is the combination of retrieval-augmented generation techniques with graphs. The (knowledge) graph is then used to ground the model by providing more accurate information than the sole use of semantic similarity. | |
| Large Action Models (LAMs) | Large Action Models (LAMs) are AI models that generate actions to achieve goals. LLMs may recommend, LAMs can effectively act. Combining both lets users express goals in natural language, with LAMs executing tasks via AI agents. LAMs revolutionize UI/UX with intuitive and efficient interactions. | |
| Legacy Code & AI | Legacy code is a problem for many organisations because of its high maintenance costs. With the advent of LLMs and coding assistants, we must find out if these smarter tools can be leveraged to better understand or modernize legacy projects. | |
| Knowledge Graphs | Knowledge Graphs are networks of interconnected entities and relationships representing real-world data in a structured form. They enable efficient data integration, retrieval, and reasoning, enhancing AI applications like search and recommendation systems. | |
| Native Graph ML | Graph Machine Learning applies ML techniques to graph-structured data. Many real-world systems can be modeled as graphs (like social networks), and Graph ML enables predictions and insights that traditional ML can’t capture. | |
| Speech Interfaces | Speech interfaces offer real-time, highly accurate speech recognition and speech synthesis, with robust support for multiple languages. They feature low latency and allow users to interrupt responses, ensuring a smooth, efficient, and user-friendly interaction experience. | |
| Local/Edge AI | AI processing moves onto edge devices (smartphones, IoT), reducing reliance on cloud and network, but requiring specific efforts in AI model optimization. Distributed or Federated Machine Learning algorithms also fit within this context. | |
| AI model transparency/Explainable AI | Any AI system deployed in the public sector has requirements regarding robustness and explainability of its outputs. To this end we investigate existing and upcoming tools that increase transparency of AI systems. | |
| Rules as Code | Concept from LegalTech, aiming to make laws, regulations, policies or documentation “computable”. A strong connection between a regulation to its implementation in code should facilitate updates, maintenance, analysis and research. | |
| Federated Learning | Federated Learning is a privacy-preserving machine learning technique that allows for collaborative model training on decentralized devices. The model is trained locally on each device where the data are stored, only the training results are sent to the central server to update the main model. | |
| Model editing | The knowledge captured in large language models can be rapidly outdated that is why LLMs need frequent retraining. Model editing is a way to update models without extensive retraining. This way, new information can be added in a very targeted way and undesirable behaviour can be modified. | |
| Self-Supervised Learning for Fraud Detection | Approach to machine learning (ML) where labels are created from the data itself, without having to rely on historical outcome data or external (human) supervisors that provide labels or feedback. Modern fraud detection systems use a combination of selfsupervised and supervised learning. | |
| Neuro-symbolic AI | Neuro-symbolic AI is the composition of symbolic AI and neural networks to leverage the strengths of both techniques while mitigating their limitations. | |
| Multimodal AI | Multimodal AI refers to artificial intelligence systems that can process and understand multiple types of data, such as text, images, audio, and video, simultaneously. This enables more comprehensive and context-aware interactions. | |
| Humanoid Robots | Humanoid robots are robots designed to resemble and mimic human behaviour and appearance. They can perform tasks like walking, talking, and interacting with humans, often used in research, healthcare, and customer service. | |
| Brain-Computer/Machine Interface | Brain computer interface is a technology that connects the brain with a machine that can be controlled with thoughts. Neural signals are captured by sensors hardware and instructions are derived from the signals. |