Sage searches gravitational-wave detector data for compact binary coalescence signals with a complete, end-to-end machine-learning pipeline. It automates data access and preparation from GWOSC, simulates realistic detector noise, generates and projects waveforms across multiple detectors, and applies signal-processing steps such as whitening, multirate sampling, and multibanding. The software trains neural networks on-the-fly with batch-wise waveform and noise generation, employs hard-noise mining strategies to improve performance at low false-alarm rates, and mitigates multiple supervised-learning biases that affect detection and generalization. Its modular design, with interchangeable model components and configurable presets, allows adaptation to new architectures, parameter spaces, and observing runs, and includes tutorials and documentation to support rapid experimentation and use.