wanted a self hosted version of the bookmarking app i already use, on the theory that the real thing is closed and a copy can be better. the pitch was specific: save something from the browser or the phone and have it appear instantly, pull the dominant color out of every image, write a short vision model explanation of what a post actually is, filter by source, and find similar saved posts by theme, style, or vibe without using a single keyword search.
the great part was the data. there was already a giant library of my saved posts on disk with clip embeddings already computed for a big chunk of them, so no re embedding and no expensive reprocessing. the new app imported the whole thing and every item became searchable. the front end got a dark mauve skin, a masonry wall of image only cards, filters tucked behind an icon, and a random browse so the feed reads like a shuffled deck instead of a paged grid. those parts went quick.
the medium part was also the most frustrating. the media pipeline pretended to work. rows saved, jobs completed, and the actual image file never existed anywhere, with no error anywhere. the round ended with the weirdest answer, a duplicate block in the pipeline, nested completely inside the wrong if two characters deep, so every normal save slipped past the download step entirely. rewriting the whole module clean fixed it in one shot.
another round existed for the design pass: the detail view needed the image left and the text right, and there was a loading flash where every card painted at the top left corner before the images streamed in. fixing that one was about laying out the grid from the known aspect ratio before the images loaded, so the slots are decided first and the media falls into place. the library ended up with a bunch of near identical duplicates collapsed away. it now behaves like the commercial tool for the parts i actually care about, and mine does not close a tab or make me feel like my data belongs to someone else.