calabrenzio/creative-autopilot-tribe

TRIBE v2 Inference Service

Small HTTP wrapper for facebook/tribev2.

The Next.js app calls this service through TRIBE_INFERENCE_URL. The service accepts a generated video URL plus prompt text/caption, runs TRIBE v2, and returns normalized attention/retention proxy scores.

Why This Exists

facebook/tribev2 is available on Hugging Face, but its README shows Python package usage:

from tribev2 import TribeModel

model = TribeModel.from_pretrained("facebook/tribev2")
df = model.get_events_dataframe(video_path="path/to/video.mp4")
preds, segments = model.predict(events=df)

That means the product needs a custom Python runtime rather than assuming a generic Hugging Face serverless Inference API.

Deploy Remotely

Do not run TRIBE v2 locally for production validation. The model and feature extractors are storage-heavy and slow on a laptop. Deploy this service to a remote runtime with persistent model cache and enough CPU/GPU resources.

Good options:

• —Hugging Face Space using the included Dockerfile
• —RunPod / Modal / Replicate / Fly machine with a persistent volume
• —A private GPU VM with the Hugging Face cache mounted on durable storage

Set the app to the deployed endpoint:

TRIBE_INFERENCE_URL=https://your-tribe-inference.example.com/score
HF_TOKEN=...
REQUIRE_TRIBE_VALIDATION=true

Local Smoke Test Only

cd services/tribe-inference
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
uvicorn app:app --host 0.0.0.0 --port 7860

Then set only for smoke testing:

TRIBE_INFERENCE_URL=http://localhost:7860/score
HF_TOKEN=...
REQUIRE_TRIBE_VALIDATION=false

Response

{
  "attention_score": 0.81,
  "retention_score": 0.74,
  "confidence": 0.68,
  "notes": ["TRIBE v2 brain-response proxy computed from predicted cortical activity."]
}

These scores are not virality predictions. They are attention/retention proxy signals that the main app combines with QA, hook quality, clarity, platform fit, and historical performance.