We improve information seeking by interactive intent modelling, and demonstrate the methods with SciNet, an interactive search system for scientific documents. The user directs the search by giving feedback for estimated search intents. In the current version of SciNet the potential intents are represented by keywords and laid out on an ""intent radar"" display which visualizes both their estimated relevances and similarity. Probabilistic language modelling and reinforcement learning are combined with new visualization methods for search in over 50 million scientific documents, outperforming conventional searches in user experiments. This is joint work with several researchers; the first papers are Glowacka et al., IUI'13 (best paper award), Ruotsalo et al., ASIST 2013, Ruotsalo et al., CIKM 2013.