Follows pull real content
Following collection:software now directly queries that collection instead of waiting for matching items to appear by chance.
On-device intelligence
A recommendation system that learns without leaving your phone.
Found's recommender is built for messy, beautiful archive data: books, films, audio, software, web captures, collections, creators, and tags. We improved it into a hybrid local system that learns from real intent while keeping your taste profile on your device.
01
Candidate Pools
We fetch a richer set of Archive.org candidates from broad media slices, learned terms, regional hints, and followed collections.
02
Hybrid Scoring
Each item is scored with metadata quality, learned behavior, saved-item similarity, and an exploration bonus.
03
Diverse Results
The top results are reranked so one creator, collection, or media type does not crowd out everything else.
The Scoring Mix
The recommender balances reliable metadata with personal behavior. It can work on a fresh install, then gets sharper as you save, follow, play, skip, and dislike items.
Score = 0.42 * Symbolic
+ 0.46 * Learned
+ 0.12 * Saved Similarity
+ ExplorationWhat We Made Stronger
Saves shape taste
Saved items act as anchors. Candidates similar to what you saved get a meaningful boost.
Passive views stay quiet
Impressions no longer train your taste. They only help avoid showing the same item too often.
Confidence matters
A single accidental tap has less power. Signals become stronger only as the app sees repeated evidence.
Intent Signals
The system treats explicit actions as stronger evidence than casual browsing. That keeps the feed personal without becoming twitchy.
SaveHighest positive intent+2.8
Play completedDeep interest+1.8
Play startedStrong interest+0.9
Open detailMild interest+0.45
SkipNegative signal-0.65
DislikeHard negative signal-3.0
Private by design
The learned profile is stored locally as on-device state. Found does not need a cloud account or remote personalization server to learn what you like.