tripkit

tripkit is a small web page that turns a trip photo into a post kit: a title, the place, a one-sentence caption and at least three hashtags, in Thai by default with an English switch, ready to paste into your own social account. A local Gemma 3 12B model does the writing through Ollama, so the photo stays on your laptop. Photos are rotated, shrunk to 1024 px and re-encoded without EXIF before the model sees them, so GPS data never reaches the model or the evaluation log.

The design rule that matters most is what the model is not allowed to do: the Place on the kit is whatever you typed, passed through unchanged, and empty with a warning if you typed nothing. An optional "let the AI guess" button only suggests a place, clearly marked as a guess that may be wrong; it never fills the Place box by itself and never reaches the title, caption or hashtags until you press "use this". Every result is checked in code for at least three hashtags and a single-line caption of at most 120 characters.

Vibe-coded evidence, in the author's own words on DEV (14 October 2026): "I built this with Claude Code. It wrote a large share of the code and the first drafts of this post, so the post is marked as AI-assisted. I made the product decisions." GitHub stars: 0 at the 28 October 2026 check (github.com/ninefyi/hf26challenge-week2).

Honest limits, stated by the author and kept in this draft: there is no hosted demo — the demoUrl here points to the repository and local run instructions (uv + Ollama, about 8 GB of memory; the page runs on localhost:7860), not to a hosted app, and the main agent should keep that disclosure if it publishes. The author also reports the place guess was clearly right only once in six public-photo trials and wrong with a "high" confidence label three times, which is why the page never shows confidence; the 70%-accepted evaluation bar was cut to a single self-judged photo and is reported as NOT TESTED, and the author does not know whether anyone would pay for it.