From not disturbing a focused programmer, to entertaining a restless commuter waiting for a train, personal ubiquitous computing devices could greatly enhance their interaction with humans, should they only be aware of the user’s cognitive load. While mobile sensing and machine learning lead to impressive advances in the inference of human movement, physical activity, routines, and other behavioural aspects, inferring cognitive load remains challenging due to a subtle manifestation of a user's mental engagement via vital signal reactions. These signals are often captured with obtrusive, expensive, purpose-built equipment, preventing seamless cognitive load inference for human - ubiquitous computing interaction adaptation. In our work we aim to enable large-scale unobtrusive cognitive load inference. In the talk I will present our experiences from different user studies in which we built and evaluated cognitive load inference models relying on data coming from a commodity smart phone, a wearable sensing device, and a software-defined-radio-based wireless radar. Finally, I will present our guidelines for future efforts in cognitive load inference and argue for closer interdisciplinary collaboration in this exciting research domain.