Shelf Scanner
An AI-powered bookshelf scanning and recommendation application. Users photograph a bookshelf, and the system automatically identifies book titles and recommends books based on the user's reading preferences. The application combines computer vision models with language models to perform title recognition and personalized recommendation.
Timeline (last 90 days)
In the Shelf Scanner evaluation, Claude Sonnet and Gemini Flash could read almost all identifiable book titles without hallucinating, while smaller models missed more titles and hallucinated more.
In the shelf-reading evaluation, latency was the primary issue, with some calls taking 10–20 seconds just to read a shelf, against an overall experience target of approximately 15 seconds.
In the recommendation stage, the cheaper GPT model actually aligned more closely with the blogger's own choices than the more expensive model.
The blogger discovered that the source images' metadata contained GPS information, which constituted a privacy issue they did not want to store.
All Facts (4)
In the Shelf Scanner evaluation, Claude Sonnet and Gemini Flash could read almost all identifiable book titles without hallucinating, while smaller models missed more titles and hallucinated more.
50%UnverifiedIn the shelf-reading evaluation, latency was the primary issue, with some calls taking 10–20 seconds just to read a shelf, against an overall experience target of approximately 15 seconds.
50%UnverifiedIn the recommendation stage, the cheaper GPT model actually aligned more closely with the blogger's own choices than the more expensive model.
50%UnverifiedThe blogger discovered that the source images' metadata contained GPS information, which constituted a privacy issue they did not want to store.
50%