An accelerated need to process information and abstract features at the edge is pushing researchers to rethink how architectures can accelerate these workloads. Conventional design trade-offs for such accelerators are either between performance and accuracy or between energy and accuracy. In this work, we propose a novel architecture where the energy efficiency varies as a function of performance at a given accuracy level and allows for workloads to be accelerated at varying degrees based on the energy available to the edge device. This is achieved through the utilization of Asynchronous Stochastic Computing (ASC) principles, wherein we implement a truly clockless, asynchronous stream-based computing paradigm, and information is encoded in the time domain by the instantaneous frequency and duty cycle of the stream. We introduce the concept of \textit{stimulus-driven workload execution}, which allows us to dynamically alter the computational rate of an accelerator based on the presence of actionable information. To validate these concepts, an event-driven image classifier is designed, and our evaluation demonstrates a scalable energy efficiency between 81.54 and 247.08 TOPS/W while maintaining an average accuracy of 98.60%. We demonstrate our solution for an MNIST dataset, but the same general principles apply to different, more complex models and datasets.