At-home health monitoring technologies have extensive use-cases, including gait assessment, disease diagnosis, gait anomaly detection and disease severity quantification. Despite the volume of research into this domain, novel machine learning models and data collection pipelines are often benchmarked on curated datasets, collected in laboratory conditions and annotated by specialists. This separation between the experimental and practical domain is apparent in the lack of interdisciplinary work in the physical design of these systems. In this work, a series of interviews with 10 healthcare professionals is conducted to co-design an at-home gait monitoring system, with the initial prototype informed by 3 focus groups of older adults. A series of recording experiments are then conducted with the prototype in the homes of 10 older adults to investigate its effectiveness and feasibility when subjected to the conditions of genuine at-home data recording. With novel ST-GCN based models achieving up to 0.93 f1 on the task of gait recognition using only data collected at home, it can be asserted that the developed gait monitor prototype is capable of reliably collecting data of sufficient quality to build accurate profiles of gait. This is achieved despite the design concessions made in the interest of stakeholder acceptability and the challenges presented by at-home gait data.