Visual programming tools have recently been introduced to enable Deep Learning (DL) development without the need for expertise in traditional programming languages and frameworks. However, these tools often exhibit limitations in scalability for complex architectures and lack real-time debugging capabilities. This paper introduces DeepFlow, a flow-based visual programming tool (VPT) designed to address these challenges by leveraging the inherently visual nature of DL models as sequences of learnable functions. DeepFlow incorporates hierarchical abstraction mechanisms through “supernodes” to support model scalability, which is crucial for modern, complex architectures. Additionally, it introduces interactive debugging in the model design phase, allowing developers to validate network architectures before execution. We conducted a user study with 16 DL developers, involving typical DL model design tasks. We assessed DeepFlow using quantitative usability metrics, and post-task interviews to evaluate user perceptions and workflow integration across different expertise levels. Results demonstrated high usability and user satisfaction, and highlighted DeepFlow’s effectiveness for rapid model iteration and as a learning aid for complex DL architectures, while also identifying areas for improvement.