Frank Vanhoenshoven - Teaching the Unknowable - How to Build an AI Preference Model with Limited Data

Amidst the ubiquitous conversations on the successes and faillures of Large Language Models and Agentic AI, it is easy to forget that their underlying models can also drive powerful purpose-built solutions that do not hallucinate, misinform or display a lack of awareness on their own pseudo-general intelligence.

In this talk, we present our contribution to a novel digital assistant developed by Trickle.ai, designed to make meeting planning less cumbersome. For this project, we built a data-driven solution capable of examining meeting timeslots and assigning them preference scores aligned with user preferences.

By walking the audience through the design process and its challenges, we highlight the incredible potential of neural networks. We present a blueprint that could easily be extended to other purpose-build solutions. The attentive attendee will go home in the absolute certainty that they too can train a Neural Networks to predict the unknowable.

Frank Vanhoenshoven and Benoît Depaire are members of BIARU, the applied research unit within the Business Informatics research group. Our mission is to bridge the gap between academic research and real-world business challenges, ensuring a two-way exchange of knowledge and experience. Over the past few years, BIARU has collaborated with multiple organizations to help them harness their data and develop solutions to concrete business problems. While we often do this by building prototype models, we also co-create production-ready solutions and deliver tailored, on-demand training activities.