This paper introduces JUICER, a data-efficient imitation learning (IL) pipeline for robotic assembly tasks that requires precise, long-horizon manipulation. The approach combines expressive policy architectures with techniques for dataset expansion and simulation-based data augmentation to improve performance with a small number of human demonstrations. The pipeline includes methods for synthetic data augmentation around bottleneck states, where small imprecisions can lead to task failure, and a "collect-and-infer" strategy that iteratively expands the dataset by incorporating successful rollouts. The system is evaluated on four furniture assembly tasks from the FurnitureBench benchmark, where it achieves high success rates by directly learning from RGB images. The results show that JUICER outperforms baselines, including imitation learning and data augmentation methods, and can achieve high-performance policies with as few as 10 human demonstrations. The paper also presents tools for bottleneck state labeling, trajectory augmentation, and a dataset of collected demonstrations for the research community. The approach is effective in both simulation and real-world settings, and the results demonstrate the potential of data-efficient imitation learning for complex robotic assembly tasks.This paper introduces JUICER, a data-efficient imitation learning (IL) pipeline for robotic assembly tasks that requires precise, long-horizon manipulation. The approach combines expressive policy architectures with techniques for dataset expansion and simulation-based data augmentation to improve performance with a small number of human demonstrations. The pipeline includes methods for synthetic data augmentation around bottleneck states, where small imprecisions can lead to task failure, and a "collect-and-infer" strategy that iteratively expands the dataset by incorporating successful rollouts. The system is evaluated on four furniture assembly tasks from the FurnitureBench benchmark, where it achieves high success rates by directly learning from RGB images. The results show that JUICER outperforms baselines, including imitation learning and data augmentation methods, and can achieve high-performance policies with as few as 10 human demonstrations. The paper also presents tools for bottleneck state labeling, trajectory augmentation, and a dataset of collected demonstrations for the research community. The approach is effective in both simulation and real-world settings, and the results demonstrate the potential of data-efficient imitation learning for complex robotic assembly tasks.