Katch · 2021 to present
Katch Party
Tinder for movies. Rate posters, get recommendations, and resolve a group's competing taste into a single pick everybody can live with.
Role · Co-founder and CTO. Product and modeling.
VisitThe problem
Two people open a streaming app and spend twenty minutes not choosing anything. The standard recommender is no help, because it models one viewer at a time and asks for behavior it does not have yet on a new account. It also asks the wrong question: people do not browse a catalog, they react to artwork.
Approach
- 01
Built preference elicitation as a poster swipe. The artwork is the stimulus, which gets a usable taste signal out of a brand-new user in under a minute and matches how people actually scan a catalog.
- 02
Mapped swipe responses onto the genome's attribute space rather than onto titles, so a few dozen reactions generalize to the long tail instead of overfitting to whatever posters happened to come up.
- 03
Added a group mode that combines several profiles into one ranked list, optimizing for the pick with the least objection rather than the highest average score, since the failure mode of a group recommender is one person hating it.
- 04
Kept the loop short. Rate, see the list, rate again, with the ranking updating in session.
Outcome
Live consumer app, and the clearest demonstration that the genome works with no interaction history behind it.