Cortex

Inspiration

Cortex was inspired by Percept, which used Meta's TRIBE v2 as a predicted brain-engagement signal for video ads. Cortex aims to turn that signal into a human-in-the-loop editing workflow.

What it does

Cortex predicts where a video ad loses a viewer's attention and allows the user to replace that 5-second moment with an AI-generated take. The new take is scored on the same brain-engagement scale against the original.

How we built it

Brain-response prediction: Meta TRIBE v2 running on Kaggle GPU T4 notebooks.
Backend: Python, FastAPI, and Pydantic.
Frontend: React 18, TypeScript, and Three.js (for the 3D brain).
Creative direction: Powered by Gemini (gemini-2.5-flash) to write positive/negative prompts.
Take generation: ElevenLabs. ## Challenges Scoring the new takes was a major challenge. Cortex uses "Same-scale take scoring" to re-normalize the AI take's score against the original ad's scale so they can be compared directly.