I need an end-to-end system that learns to create eye-catching YouTube thumbnails from scratch. The core will be a computer-vision model that looks at colours, contrast, facial expressions and any headline text or fonts I feed it, then composes a thumbnail aimed at maximising click-through rate. Because I’m starting with no historic dataset, part of the job is designing a training pipeline that can bootstrap its own learning—first with publicly available video frames and titles, later by ingesting live performance data so the model keeps refining itself. The workflow I have in mind is straightforward: I hand the service a video (or individual frames plus a title), it returns a ready-to-upload PNG/JPEG sized for YouTube, plus the underlying metadata so I can track which visual decisions drove results. Python with PyTorch or TensorFlow, OpenCV and a lightweight web API feel right, but I’m open to alternatives if they speed things up or improve accuracy. Deliverables I can easily plug into my channel management routine: • A trained, self-updating thumbnail generator accessible through a REST or CLI interface • All training scripts, model weights and a short README for retraining or fine-tuning • Simple logging that records CTR feedback so the system can adapt over time The build should run on a standard GPU instance (local or cloud) and generate its first batch of thumbnails before we wrap.