BrainCurve
Turning a video into a second-by-second map of attention, emotion, and memory
THE
PROBLEM
Creators and researchers can measure whether a video performed, but not why — which exact moments hold attention, land emotionally, or are likely to be remembered. Existing analytics are aggregate and after-the-fact; there was no way to see engagement as a continuous signal over the timeline.
OUR
APPROACH
BrainCurve builds on Meta’s TRIBE multimodal brain-encoding model: audio (HuBERT), vision (DINOv2), and text embeddings are fused to predict per-timestep cortical responses, then distilled into attention, emotion, memory, and visual curves for any video. A FastAPI service handles upload, feature extraction, and inference; a Next.js frontend surfaces the engagement timeline, an overall score, and the strongest and weakest moments with explanations.
Engagement timeline — attention, emotion, memory, and visual curves over the video
Per-metric neuro-scores with methodology
Key moments and a per-video breakdown
IMPACT
PROCESS
RESEARCH
Studied TRIBE and multimodal brain encoding to map the approach to a usable product
FEATURE PIPELINE
Built extraction with HuBERT (audio) and DINOv2 (vision) plus text embeddings
INFERENCE ENGINE
Ran TRIBE to predict per-timestep responses, with GPU/CPU device fallback
DISTILLATION
Turned cortical predictions into attention, emotion, memory, and visual curves
PRODUCT
Wrapped it in a FastAPI service and a Next.js dashboard with key-moment explanations
TECHNOLOGIES & SKILLS
