Research
At Traize, we research model post-training, information retrieval and LLM reasoning. Our goal is to make AI more capable, efficient and affordable for the companies using it.
We work on how models find information, use tools and solve tasks with company knowledge. This means evaluating models for the work they need to do, improving the systems around them and working towards models trained specifically for these tasks. We want every improvement to translate into better answers, faster workflows and lower costs for your team.
Here, our researchers share their take on recent papers in generative AI: what the findings mean, where they could be useful in a business and how they inform our work on Traize. You can also explore selected publications from our team.
Research perspectives
Our researchers examine new papers in generative AI, explain the ideas behind them and assess where they could make a practical difference. We share what we find promising and how it shapes our approach at Traize.
Article · 2 min read
A passage becomes easier to find when the search system understands where it belongs. Research on contextual embeddings explores how to preserve that information efficiently.
Article · 3 min read
A new benchmark compares ten LLMs with 26 embedding models across 37 tasks. The results show why companies should decide carefully which work a language model does.
Article · 2 min read
Business workflows contain many recurring, well-defined tasks. Specialised models could make these steps cheaper to run and easier to control.
Article · 2 min read
Finding a related document is only the beginning. Research on specialised rerankers shows how compact models can learn to judge whether a passage actually helps answer a question.
Our research
Our research spans how models learn, how agents make decisions and how reliable systems scale. These publications offer a closer look at the expertise we bring to Traize.
Our researchers
Our team brings together research experience in machine learning, language models and distributed systems. These are the people developing the ideas behind Traize.
Information retrieval, knowledge graphs
Leads the technical architecture. As an M.Sc. student in Artificial Intelligence in Amsterdam and with several years of experience in machine learning engineering, he builds the retrieval algorithms that power the Traize knowledge graph.
Generative AI, language model integration
Leads the integration of the language models. With an M.Sc. in Bioinformatics from TUM, he ensures generative AI works reliably and transparently in an enterprise context.
Security and data sovereignty
Drives security and data sovereignty. Designs the architecture that protects sensitive corporate knowledge and ensures its availability, from the encryption concept to access control.
Research focus
We focus on three areas that shape the quality and cost of AI at work.
Company tasks require models to recognise relevant information, use tools correctly and follow instructions reliably. We investigate how targeted training can adapt pretrained models to these demands. In the longer term, we want to offer models trained specifically for tasks in Traize.
The answer to a question is often spread across documents, tasks and conversations. We investigate which structures and relationships can be prepared as information arrives, so agents can quickly find the right context later. Our goal is more precise retrieval and less repeated processing per query.
An agent needs to decide which source to check, which tool to call and when it has enough evidence for an answer. We investigate how to guide and evaluate these decisions. Our aim is to solve complex tasks with fewer unnecessary steps, improving answer quality, computing costs and speed together.
Companies relying on proprietary model APIs are exposed to changes in pricing, availability and model behaviour. General-purpose models also need to navigate company knowledge, follow access rules and choose the right tools across complex tasks. We investigate how targeted post-training of European models that companies can deploy and operate under their own control can improve these capabilities while reducing provider dependence and keeping training and operating costs manageable.
Explore what Traize can do with your company’s knowledge. For research questions or collaboration, get in touch with our team.