This paper presents the development of a classifier for real-time flute Instrumental Playing Technique (IPT) recognition. Our classifier, implemented using Convolutional Neural Networks (CNN), shows a solid state-of-the-art accuracy in detecting different flute IPTs, with enough reactivity and steadiness to be accurate in real-time applications. In particular, we use our classifier's output to guide Somax2, a co-creative Corpus-Based Concatenative Synthesis (CBCS) improvisation system developed within the ERC project REACH, of which we are part of. By implementing a new Label Feature in the corpus annotation and in the output selection process of Somax2, we can drive its musical generation through the flute IPT detected by the classifier. Our real-time IPT recognition system offers a new dimension to the Somax2 interaction paradigm, where its artificial agents are now able to responsively engage with the recognized techniques. Contributing to the broader field of human-machine interaction in computer music, our results have potential applications in improvisation, computer-aided composition and new interfaces for musical expression.