In this paper, we present a pruning and fine-tuning framework for compressing BERT for deployment on resource constrained hardware. Our method identifies a three-step process to achieve automatic pruning on BERTBASE, and deployment of this compressed model onto the popular edge device Raspberry Pi 5. We evaluate the approach on GLUE tasks (SST-2), demonstrating strong pre and post-fine-tuning accuracies compared to BERTTickets. We provide a theoretical hardware analysis, estimating FPGA speedups relative to CPU and GPU, and show a 30% sparsity leads to a 1.43x theoretical latency improvement on a Xilinx Alveo U200. Our edge deployment on a Raspberry Pi 5 shows how resource constrained hardware compares with larger CPUs and GPUs. These results highlight high viability on pruning driven compression for edge deployments. Transformer-based models have achieved remarkable suc- cess across a wide range of natural language processing (NLP) tasks, including sentiment classification [8], question answering [9], text summarization [10], and language model- ing [11]. Among them, BERT [7] has demonstrated substantial improvements on the GLUE benchmark [12], a widely adopted suite for evaluating language understanding systems. As deep learning becomes increasingly part of industry [13], there is growing interest in deploying BERT on resource-constrained devices that require low-latency and high-accuracy inference. Prior work has explored identifying sub-networks of the BERTBASE model with desirable sparsity and accuracy [14]– [16], but these approaches rely on expert-designed heuristics that are often time-consuming and fail to yield globally optimal solutions [17]. Automatic weight is computationally expensive, requiring repeated sub-network sampling and multi- epoch retraining. Moreover, pruning Transformer models re- mains particularly challenging due to the sensitivity of syntax and semantics in language tasks. Motivated by these limitations, we introduce a simple prun- ing framework for BERT that samples candidate sub-networks, evaluates them without fine-tuning, and subsequently fine- tunes the most promising candidates. Our approach aims to identify high-quality sub-networks without requiring expert in- tervention and eliminates the need for the thousands of training epochs typical of existing automatic pruning methods