Continuously tracking cognitive demands via pupil dilation is a desirable goal for the monitoring and investigation of cognitive performance in applied settings where the exact time point of mental engagement in a task is often unknown. Yet, hitherto no experimentally validated algorithm exists for continuously estimating cognitive demands based on pupil size. Here, we evaluated the performance of a continuously operating algorithm that is agnostic of the onset of the stimuli and derives them by way of retrospectively modeling attentional pulses (i.e., onsets of processing). We compared the performance of this algorithm to a standard analysis of stimulus-locked pupil data. The pupil data were obtained while participants performed visual search (VS) and visual working memory (VWM) tasks with varying cognitive demands. In Experiment 1, VS was performed during the retention interval of the VWM task to assess interactive effects between search and memory load on pupil dilation. In Experiment 2, the tasks were performed separately. The results of the stimulus-locked pupil data demonstrated reliable increases in pupil dilation due to high VWM load. VS difficulty only affected pupil dilation when simultaneous memory demands were low. In the single task condition, increased VS difficulty resulted in increased pupil dilation. Importantly, online modeling of pupil responses was successful on three points. First, there was good correspondence between the modeled and stimulus locked pupil dilations. Second, stimulus onsets could be approximated from the derived attentional pulses to a reasonable extent. Third, cognitive demands could be classified above chance level from the modeled pupil traces in both tasks.