Thumbnail Optimizer

Content creators routinely pick a thumbnail on instinct, with no structured way to compare candidates before publishing and no feedback loop until the content is already live and switching has little upside. Thumbnail Optimizer is meant to replace that guess with a scored, explainable comparison, backed up by a real A/B test rather than blind trust in the score.

The plan: a creator uploads two or more candidate thumbnails, a computer-vision model scores each for predicted click-through rate and explains the score in plain language (referencing things like contrast, face presence, or text legibility), and the creator can either act on the ranking directly or launch a structured A/B test to confirm it against real engagement data. None of that is built yet — right now the project only has its repo scaffold, authentication, and core data models in place.

The target architecture is a React + Tailwind dashboard talking to a Node.js/Express API, which hands image-scoring and explanation work off to a separate Python FastAPI service (OpenCV for feature extraction, LangChain for generating the plain-language explanation), with MongoDB holding users, thumbnails, predictions, and A/B tests. That's the design in the project's PRD, not yet the implementation — the CV scoring engine and A/B testing module are both still unbuilt.

None yet. This is a scaffold-stage project — auth and data models exist, but the parts that would produce a result (the CV scoring engine and the A/B testing module) haven't been built.

Build the CV scoring/explanation engine first, then the A/B testing module, in that order — those are the two pieces the rest of the product depends on.