Blindspot_Meta

A hard-looking image is not necessarily a trustworthy AI test. BlindSpot helps developers distinguish model extraction errors from problems introduced by image generation or AI review. This demo shows actual synthetic invoice photographs, rejected inputs, recorded Muse Spark results and a separate Wonder-powered scenario-design and coaching workflow. Built with: Muse Wonder — scenario design Muse Image — visual generation and photographic edits Muse Spark — extraction, inspection and prompting feedback TypeScript, React, Express and Sharp Featured result: Five photographic invoices and three original controls passed the eight-field extraction task. No extraction weakness was found in that pilot. Generated-content drift and inaccurate reviewer measurements are shown separately. Limitations: Synthetic inputs, assistant-reviewed references and a narrow extraction task—not full OCR accuracy or general robustness. Invoice execution and autonomous receipt coaching are separate workflows. This video uses synthetic narration and recorded experiment evidence, not fresh live API execution.