This paper reports on the ongoing international project "System architecture for ICALL" and the progress made by the Swedish partner. The Swedish team is developing a web-based exercise generator reusing available annotated corpora and lexical resources. Apart from the technical issues like implementation of the user interface and the underlying processing machinery, a number of interesting pedagogical questions need to be solved, e.g., adapting learner-oriented exercises to proficiency levels; selecting authentic examples of an appropriate difficulty level; automatically ranking corpus examples by their quality; providing feedback to the learner; and selecting vocabulary for training domain-specific, academic or general-purpose vocabulary. In this paper we describe what has been done so far, mention the exercise types that can be generated at the moment as well as describe the tasks left for the future. [For the complete volume, see ED574893.]