Project: GETTR Vision Recommender System
Role: System architect, ML pipeline lead
Problem: Static, one-size-fits-all short video feed; no metadata, poor personalization
Solution:
Extracted keyframes → classified with CNNs + CLIP
Captioned with BLIP → generated semantic tags
Recommender engine built on content similarity using vector search
Impact:
2x increase in daily video views
35%+ boost in average watch session
Better discovery for new content and diverse creators
In 2022, GETTR launched GETTR Vision—a short-form video feature built to ride the wave of TikTok-style content. It had the right ingredients: political voices, global creators, and viral potential. But we quickly hit a wall that every social video platform eventually faces:
People weren’t seeing videos they actually cared about.
The system was live. The content was growing. But user engagement plateaued. Retention started to slide. The problem? Recommendation.
At launch, we took the pragmatic path: recommend based on recency, popularity, and watch time.
It was a basic heuristic system: