Piece ittogether.
Photos, timestamps and places scattered across your phone. MemoryShards pieces them back into the order they happened.
No manual albums.
No tagging.
Just photos in. A day out.
One fragment at a time
This is what one reconstructed fragment looks like — the same thing this project does across an entire timeline.
How the pieces get found.
Five stages. Each one shows the bar it has to clear.
Extract
Reads time and GPS out of the photo file itself.
Cluster
Groups photos into events by time and place.
Recognize
Finds every face and groups them by person.
Caption
Describes each event in a sentence.
Reconstruct
Fuses events, places and people into one timeline.
What comes out the other side.
The example this project was scoped around — a day scattered across four unrelated fragments, reassembled.
Measured on faces it never saw.
Face verification accuracy, up from 59% before we fine-tuned it. 200 pairs from 29 held-out people.
Pairwise precision grouping 230 faces by person. Recall is 76%: it splits a person before it merges two.
Model trained by us, out of five stages. The rest is classical logic or a pretrained model used as-is.
The limits: children are better but not solved, the same person decades apart is often split in two, and recall is lower on real phone photos.
Now try it on your own photos.
Upload a few real photos and watch the actual pipeline run — EXIF extraction, reverse geocoding, event clustering, and captioning — live, on a dedicated page.
Try it on your own photos →Still turning over
The same four fragments, no longer scattered — just frames in one reel now.