Built a reproducible DICOM-to-triage workflow on synthetic CT-like data. Validates and assembles DICOM series into 3D volumes, trains a custom PyTorch 3D CNN with MONAI Dice + BCE loss, and benchmarks it against a threshold baseline on held-out data. Predictions feed an auditable SQLite reading queue with model SHA-256 provenance and an HTML dashboard. Research demo only, not a diagnostic device.