AI Agent · 2025
ScrollMind
Controllable video recommendation: natural-language intent overrides the algorithmic feed (multimodal understanding + LLM rule generation + re-ranking)
Case Study
A controllable video recommendation system that allows users to express explicit viewing intent via natural language, dynamically overriding the default algorithmic feed.
Methods
Multimodal content understanding: video thumbnails + text features → 11-category classification with confidence scores
LLM-powered rule generation: natural language input → structured filtering rules (time slot + content type)
Intent filter injected post-recall, soft-weighting candidates before ranking
Re-ranking stage further adjusted via intent embedding to surface intent-matching content
Pipeline
Future Work
Extend to implicit intent modeling — infer user mood/context from watch behavior rather than requiring explicit input
