Skip to main content
Category: GTM Intelligence · Version: 1.0.0 · Author: Varnan-Tech · Compatibility: All targets (claude, codex, opencode, gemini, anti-gravity, openclaw, hermes)
AI Skill that uses Meta’s TRIBE v2 fMRI Model to analyze the neuroscience of video hooks, reels, and scripts. Instead of guessing what makes a hook engaging using prompt engineering, this skill predicts actual human brain activity across the scientifically validated Yeo-7 Functional Networks, giving you an evidence-based Engagement Report for your content. meta-tribeV2-skill-cover-image

​
Install

​
What This Skill Does

This skill provides the infrastructure to host the massive 80GB TRIBE v2 model pipeline and gives your AI Agent the ability to:
  1. Process video, audio, or text scripts.
  2. Intercept and optimize the media (downscaling video to 360p at 10fps to avoid hour-long bottleneck processing times).
  3. Process the content through V-JEPA (Vision), W2V-BERT (Acoustics), and LLaMA 3.2 3B (Linguistics).
  4. Predict human brain fMRI activity across the Yeo-7 networks.
  5. Generate an actionable, human-readable neuroscience report without the jargon.

​
Deployment Options

Because TRIBE v2 requires a massive amount of VRAM (24GB for text, up to 80GB for video), we offer 3 different deployment options so anyone can use it, regardless of budget or technical expertise.

​
1. Google Colab (Zero Cost, Decoupled)

Best for users without a cloud budget. Colab provides free T4 GPUs.
  • How it works: We use a decoupled architecture. You run the heavy AI inference in a Colab Notebook, which outputs a preds.npy prediction file. You then run a local script on your laptop to generate the report.
  • Setup:
    1. Open Google Colab and upload the script from scripts/colab_inference.py into a new Notebook.
    2. Run the notebook. It will output preds.npy and segments.json.
    3. Download those files to your machine and run: python scripts/local_analyze.py --preds preds.npy. This will output a text report and an ASCII terminal graph showing the engagement peaks and valleys.

​
2. RunPod (Serverless, Pay-per-second)

Best for production agents and developers. You only pay for the seconds the model is running.
  • How it works: We provide a RunPod Handler and a custom Dockerfile that caches the 80GB model inside the image.
  • Setup:
    1. Build the Docker image using server/Dockerfile.runpod: docker build -f Dockerfile.runpod -t tribe-runpod .
    2. Push the image to Docker Hub or GHCR.
    3. Create a new RunPod Serverless Endpoint using your image URL.
    4. Point your AI Agent to your RunPod Endpoint URL.

​
3. AWS EC2 Persistent (Enterprise, BYO-Compute)

Best for heavy, continuous usage.
  • How it works: Automatically provisions an AWS g5.12xlarge instance (4x A10G GPUs) and runs a FastAPI server.
  • Setup:
    1. Ensure your AWS account has a vCPU quota limit of at least 48 for “Running On-Demand G and VT instances”.
    2. Run bash scripts/launch_persistent.sh to provision the instance.
    3. Run export HF_TOKEN=“your_token” followed by bash scripts/deploy_to_persistent.sh to build and launch the Docker API.

​
AWS GPU Lifecycle & Estimated Costs

Running the g5.12xlarge instance (4x A10G GPUs) provides incredible speed but costs $7.09 per hour on On-Demand pricing. It is crucial to manage this lifecycle.
  1. Launch: Run bash scripts/launch_persistent.sh (Takes ~3 minutes).
  2. Analyze: Run your videos through the API.
  3. Terminate: When you are completely finished for the day, you MUST terminate the instance to stop billing.
    • Run aws ec2 describe-instances --filters "Name=instance-state-name,Values=running" to find your Instance ID.
    • Run aws ec2 terminate-instances --instance-ids <YOUR_INSTANCE_ID>.
    • Do not just “stop” the instance if you don’t want to pay for EBS Volume storage costs overnight. Terminate it.

​
HuggingFace Authentication (Required for all methods)

TRIBE v2 relies on meta-llama/Llama-3.2-3B, which is a Gated Model.
  1. Create a HuggingFace account.
  2. Go to the Llama 3.2 3B page and TRIBE v2 page and agree to Meta’s license terms.
  3. Generate a HuggingFace Access Token (Read permissions) at huggingface.co/settings/tokens.
  4. Supply this token via the HF_TOKEN environment variable.

​
The Neuroscience of the Engagement Report

The AI agent will read the raw API output and translate the neuroscience into plain English for you:
  • VAN (Ventral Attention Network): Translated to “Is this surprising enough to stop a scroll?”. High VAN means the content is novel and creates a pattern interrupt.
  • DMN (Default Mode Network): Translated to “Will people get bored and tune out?”. High DMN is bad. It means the brain is wandering. The AI uses this to identify “Cut Candidates” in your video.
  • DAN (Dorsal Attention Network): Translated to “Are people actively following along?”. High DAN means strong logical focus.
  • Limbic Network: Translated to “Does this make people feel something?”. High Limbic means strong emotional response.
Check out the Results Showcase for actual examples of Neuro-Marketing reports generated by this skill.