This paper presents an end-to-end approach for deploying and evaluating pedestrian crossing intention prediction on an on-board embedded computing platform. The proposed pipeline integrates YOLO11s for pedestrian detection, DeepSORT for identity tracking, RTMPose-s for pose estimation, an identity-aware temporal buffer, and PedAST-GCN for crossing intention prediction. We optimize the pipeline using ONNX Runtime and TensorRT and deploy it on an NVIDIA Jetson AGX Orin installed in KTH’s Research Concept Vehicle E. In addition, we introduce an offline LLM-based evaluation framework that automatically annotates pedestrian crossing states from real-world tracking videos and evaluates whether the system can anticipate crossing behavior one to two seconds in advance. Experiments conducted at the Gillinge test track demonstrate reliable prediction performance and real-time inference at approximately 16.6 FPS. This replication package provides the source code and documentation required to reproduce the deployment, optimization, and evaluation procedures reported in the paper.