FLdevBoo

FrameCheck answers a question every front-end developer asks: does the build match the design? The problem: developers check their work by flipping between a Figma export and the browser, and small defects slip through: a button that never rendered, a headline cut off, a card that moved 30 px. Pixel-diff tools flag every rendering and compression change and cannot say what went wrong, so reviewers stop reading them. What it does: upload a design export and a screenshot of the build. One Meta Muse Spark request compares both images and returns findings in three categories (missing element, clipped text, layout mismatch), each with a plain-language summary, a check step, and a box on the design, the build, or both. The developer reviews every finding side by side, in a swipe view or in 3D layers, accepts or rejects it with a reason, and exports the reviewed run as Markdown for a pull request or JSON for tooling. Nothing counts until a person decides. How it uses Meta AI: both images go to Muse Spark (muse-spark-1.3) in a single /v1/responses request with a strict JSON schema. The model answers on a 0–1000 grid; FrameCheck stores the exact downscale transform and maps every box back to original pixels. Output is treated as untrusted: each finding is validated against evidence rules for its category, developer notes cannot change the schema, and a request cap, duplicate-run detection and token accounting keep cost under control. In 4 live runs on images with injected defects, Muse Spark found 12 of 12 with 0 false positives (box IoU 0.92–0.96). Built with Go, PostgreSQL, React, TypeScript and the Meta Model API. MIT licensed.