AI on the Edge: An Automated Pipeline for PyTorch-to-Android Deployment and Benchmarking

The deployment of deep learning models on mobile devices is a cornerstone of modern AI applications. However, performance benchmarking in this domain remains a predominantly manual, time consuming, and non scalable process. This paper introduces a fully automated, end-to-end pipeline NN Lite \url{https://github.com/ABrain-One/nn-lite} that bridges the critical gap between model development in PyTorch and rigorous performance evaluation on the Android platform. Our system comprises a Python based orchestration framework that manages model conversion, emulator control, and data collection, working in tandem with a lightweight Android application for on device benchmarking. The orchestrator systematically converts PyTorch models to TensorFlow Lite format, deploys the benchmark application, executes inference tests, and retrieves detailed performance reports. The output is a collection of structured JSON reports containing precise inference latency metrics enriched with device specific hardware analytics. This framework eliminates manual intervention, ensures reproducibility, and provides a scalable solution for evaluating the on device performance of diverse neural network architectures. In a large scale evaluation, the system successfully processed over 7,500 models, demonstrating exceptional with 48+ hours of continuous unattended operation, thereby establishing a new standard for automated mobile ML testing infrastructure.