The problem
An AI-native podcast company produces weekly episodes 60-90 minutes long. Currently the highlights are hand-selected by producers. Build a pipeline that takes the audio + transcript, generates 3-5 candidate highlight clips (30-60 sec each), ranks them by predicted audience engagement (based on past hit clips), and produces caption-ready output. Should work from transcript alone where audio isn't available, and improve when audio IS available (prosodic features).
What success looks like
Pipeline runs in under 10 minutes per episode. Top-3 predicted highlights overlap with human-selected highlights at least 50% of the time on a 10-episode holdout set.
Data access
50+ past episodes with transcripts and historical highlight performance (views, retention) provided.
Constraints
Open-source baseline comparison required (e.g., naive segment-by-keyword extraction).
Reward
$1,500 team prize + live podcast feature