This paper presents a solution for achieving crisp interactive response in resource-intensive applications on resource-poor wearable hardware. The approach is based on multi-fidelity computation supported by predictive resource management. The system automatically makes runtime fidelity decisions on the applications' behalf, freeing programmers from this burden. It exploits history-based prediction of application resource usage to improve performance. The system, called Odyssey, is implemented on top of the Odyssey framework, which originally supported the concept of fidelity for stored data. This work extends that concept to the broader notion of computational fidelity and demonstrates its applicability to a new class of applications. The system's design is based on the notion of an operation, which is the smallest user-visible unit of execution. Each operation corresponds to one multi-fidelity computation, with fidelity metrics settable at operation start. Odyssey performs five predictive mappings to determine the optimal fidelity value for an operation. These mappings include predicting operation latency, resource demand, and user satisfaction or utility. The system's performance is evaluated using four applications, with results showing a 60% reduction in mean latency and a 30% reduction in the coefficient of variation for one application. The system also shows that history-based prediction of resource demand is feasible, accurate, and necessary for this improvement. The system's design rationale includes three fundamentally different approaches to coping with situations where application resource demand exceeds supply. The first approach is to prevent such situations by using QoS-based resource reservations. The second approach is to acquire additional resources through remote execution. The third approach is to reduce resource demand through multi-fidelity computation. The system's interface and implementation include a programming interface based on the notion of an operation. The system's architecture includes supply predictors, performance predictors, a solver, demand monitors, and a logger. The system's performance is evaluated using a benchmark of two applications: GLVU and Radiator. The results show that the system can significantly reduce mean latency and variability in latency for interactive applications. The system's costs and overheads include porting costs and runtime overheads. The system's porting costs are modest, involving 500–1000 additional lines of code per application. The system's runtime overheads are around 20 ms, which is acceptable for applications with a 1 s latency bound but unacceptable for applications with a 100 ms latency bound. The system's related work includes previous work on fidelity adaptation, QoS-based reservations, and remote execution. The system's approach is different from traditional models of adaptation by using a predictive rather than a feedback-driven approach. The system's approach is also different from previous work on resource prediction by using history-based prediction to model resource demand as a function of fidelity in adaptive applications. The system is the first to use history-based prediction to model resource demand as a function of fidelity in adaptive applications.This paper presents a solution for achieving crisp interactive response in resource-intensive applications on resource-poor wearable hardware. The approach is based on multi-fidelity computation supported by predictive resource management. The system automatically makes runtime fidelity decisions on the applications' behalf, freeing programmers from this burden. It exploits history-based prediction of application resource usage to improve performance. The system, called Odyssey, is implemented on top of the Odyssey framework, which originally supported the concept of fidelity for stored data. This work extends that concept to the broader notion of computational fidelity and demonstrates its applicability to a new class of applications. The system's design is based on the notion of an operation, which is the smallest user-visible unit of execution. Each operation corresponds to one multi-fidelity computation, with fidelity metrics settable at operation start. Odyssey performs five predictive mappings to determine the optimal fidelity value for an operation. These mappings include predicting operation latency, resource demand, and user satisfaction or utility. The system's performance is evaluated using four applications, with results showing a 60% reduction in mean latency and a 30% reduction in the coefficient of variation for one application. The system also shows that history-based prediction of resource demand is feasible, accurate, and necessary for this improvement. The system's design rationale includes three fundamentally different approaches to coping with situations where application resource demand exceeds supply. The first approach is to prevent such situations by using QoS-based resource reservations. The second approach is to acquire additional resources through remote execution. The third approach is to reduce resource demand through multi-fidelity computation. The system's interface and implementation include a programming interface based on the notion of an operation. The system's architecture includes supply predictors, performance predictors, a solver, demand monitors, and a logger. The system's performance is evaluated using a benchmark of two applications: GLVU and Radiator. The results show that the system can significantly reduce mean latency and variability in latency for interactive applications. The system's costs and overheads include porting costs and runtime overheads. The system's porting costs are modest, involving 500–1000 additional lines of code per application. The system's runtime overheads are around 20 ms, which is acceptable for applications with a 1 s latency bound but unacceptable for applications with a 100 ms latency bound. The system's related work includes previous work on fidelity adaptation, QoS-based reservations, and remote execution. The system's approach is different from traditional models of adaptation by using a predictive rather than a feedback-driven approach. The system's approach is also different from previous work on resource prediction by using history-based prediction to model resource demand as a function of fidelity in adaptive applications. The system is the first to use history-based prediction to model resource demand as a function of fidelity in adaptive applications.