The Problem
A trip's camera roll is mostly noise. You shoot the same view five times hoping one is sharp, half the frames are blurry or accidental, and afterwards there's no sense of where any of it happened. Culling it by hand is the tedious part nobody does, so the good photos drown. And mainstream photo apps don't help — they optimize for sharing and engagement, not a clean, durable record of your own days.
The good photos drown. Culling is the part nobody does.
The Approach
Trippin' is the successor to FotoLog's map-first journal, with an automatic curation pipeline doing the work you wouldn't:
- Extract. Every photo is mined for signal — a perceptual hash, orientation and device tags, the time of day (it knows golden hour), blur and saliency for free, then a caption, a focal point, and an aesthetic score from a vision model. Far more than a beauty score; it's the signal everything downstream leans on.
- Embeddings. An OpenCLIP embedding (512-d) lands in
pgvector, so the library is searchable by what's in a photo, not its filename. - Reframe. A rule-of-thirds crop positioned on the focal point — the extracted vision point when present, local saliency otherwise, centered only as a last resort.
- Dedup with judgment. Near-duplicates are clustered by pHash and the keeper is chosen by quality (aesthetic score, falling back to sharpness), not by upload order.
The trip-wide run is two-phase with a barrier: the per-photo work must reach a terminal state before the single trip-wide dedup pass runs — so deduplication compares fully-scored photos, not half-processed ones. Every step is a small, side-effect-free verb, which means an agent can inspect a photo before deciding to act on it.
And the review model is deliberately conservative: the system flags duplicates and junk but never deletes. You confirm before anything leaves.
The Outcome
A trip curates itself into a map and a timeline you'd actually revisit — the best shot of each moment surfaced, the bursts collapsed, the junk set aside, and the whole library searchable by what's in the photo rather than its filename. Nothing gets thrown away behind your back.