On-demand platforms, exemplified by companies like Uber and Lyft, are a disruptive business model. Requests (e.g., a ride, a delivery) are fulfilled by matching independent suppliers (e.g., freelance logistics providers) with demand requests. Current centralized approaches to platform design excel at meeting demand commitments, but limit supplier autonomy. Decentralized approaches provide supplier autonomy, but sacrifice systematic performance and are time consuming. This research proposes a new hierarchical approach, recasting the platform's role as one providing personalized recommendations (i.e., a menu of multiple requests) to suppliers. Supplier choice can increase participation (capacity) and resource utilization when request fulfillment is combined with suppliers' original planned tasks. However, supplier choice also makes task allocation more complicated, requiring new methods to determine which requests and how many to recommend, as well as strategies to mitigate outcomes of supplier choice (e.g., rejections, duplicate selections). We consider how to design and operate a platform that first must decide how multiple, simultaneous recommendations are made. Then, suppliers have autonomy to select requests (if any) from the personalized recommendations. This results in some requests not selected and others with duplicate selections. The platform determines recourse actions for these requests, which may be fulfilling rejected requests using platform resources or recommending them to another supplier. To guide design questions and to quantify the impact of supplier choice on platform efficiency, effectiveness, and equity, we create a bi-level optimization framework. These models are novel as they capture the interdependent outcomes of supplier selections. Able to solve large problems reasonably quickly, we use this framework to answer the open research question: when is supplier choice beneficial? Applications of this approach to crowdsourced delivery, ride sharing, and on-demand volunteer platforms are explored. This research is partially funded by the National Science Foundation award 1751801 and is joint work with Shahab Mofidi, Rosemonde Ausseil, Hannah Horner, Safron Smith, and Professor John Mitchell.