What is happening inside the viewer's brain while they watch?
Nuro is a neural attention analytics platform that reveals how content feels to viewers — second by second — using brain encoding AI.
Content creators and ad agencies measure performance with behavioral proxies: watch time, CTR, scroll depth. These answer how much, never how. Nuro answers the real question: what is the viewer's brain doing at each moment?
YouTube URL
↓
GPU Server (Modal / A100)
→ Download video via yt-dlp
→ Extract audio, video frames, transcript
→ TRIBE v2 brain encoding inference
→ Output: (n_seconds ×ばつ 20,484) cortical activation matrix
→ Compute 4 metrics, normalize 0–100
↓
Claude Opus API
→ Identify peak engagement windows, drops, spikes
→ Generate actionable insight cards per moment
↓
Next.js Dashboard (Vercel)
→ Interactive second-by-second chart
→ Metric lines with hover tooltips
→ Insight cards with timestamps + recommendations
Source: Meta FAIR (Fundamental AI Research)
Paper: "A foundation model of vision, audition, and language for in-silico neuroscience" — ICLR 2026
Weights: facebook/tribev2 on HuggingFace
TRIBE v2 is a trimodal foundation model trained on 1,000+ hours of fMRI recordings from 720 subjects. It predicts high-resolution brain responses to any audio-visual stimulus — without requiring actual brain scans.
| Input | Encoder | Brain Region |
|---|---|---|
| Video frames | V-JEPA2 | Visual cortex (V1–V4, LO) |
| Audio | Wav2Vec-BERT | Auditory cortex |
| Transcript | Llama 3.2 | Language network (Broca, STG) |
Output: (n_seconds ×ばつ 20,484) — predicted fMRI activation for every cortical vertex, every second. Every number corresponds to a real brain region with known function. This is not a proxy metric — it is a simulation of neural activity validated against decades of empirical neuroscience.
attention = preds[:, 0:500].mean(axis=1) # Visual cortex (V1–V4) emotional = preds[:, 1500:2000].mean(axis=1) # Default Mode Network cognitive = preds[:, 1000:1500].mean(axis=1) # Language network (Broca) memory = preds[:, 2000:2200].mean(axis=1) # Parahippocampal gyrus # Normalize to 0-100 normalized = (x - x.min()) / (x.max() - x.min()) * 100
| Metric | Brain Region | High = |
|---|---|---|
| Attention | Visual cortex (V1-V4) | Eyes locked, full visual engagement |
| Emotional | Default Mode Network | "This means something to me" |
| Cognitive Load | Language network (Broca) | Complex message being processed |
| Memorability | Parahippocampal gyrus | This moment will be remembered |
Next.js (Vercel)
-> /api/analyze (Next.js API route)
-> Modal GPU endpoint (FastAPI + TRIBE v2)
-> Claude Opus API (insight generation)
-> Combined JSON response to frontend
- GPU: A100 40/80GB on Modal (auto-scaling) or RunPod
- Inference time: ~5-10 min per 10-minute video
- Cost per analysis: ~0ドル.50-1.00 GPU + ~0ドル.05-0.10 Claude API -> ~98% gross margin at 75ドル/analysis
| Nuro | Neurons AI | Tobii | Focus Group | |
|---|---|---|---|---|
| Method | Brain encoding AI | Eye-tracking + EEG | Eye-tracking | Survey |
| Input | Any video URL | Static images | Video/screen | Physical session |
| Speed | Minutes | Minutes | Minutes | Weeks |
| Cost | ~75ドル | ~500ドル+ | Hardware required | 5,000ドル-50,000 |
This repo contains the GPU inference backend deployed on Modal.
Requirements: Modal account, Hugging Face token with access to facebook/tribev2
# Set up Modal secret (HuggingFace token) modal secret create hf-secret HUGGING_FACE_HUB_TOKEN=hf_your_token # Deploy cd tribev2-modal modal deploy app.py
For RunPod deployment, see RUNPOD.md.
Test the endpoint:
curl -X POST "https://your-workspace--tribe-v2.modal.run/predict" \ -F "file=@/path/to/video.mp4" \ -o predictions.zip
Python FastAPI PyTorch Modal Next.js TypeScript Tailwind CSS Claude Opus API yt-dlp
Built by Kutluhan Eniste - Istanbul, Turkey TRIBE v2 is CC-BY-NC-4.0 (Meta FAIR)