ZeroDriveX / Meta AI Hackathon — Startup Track

ZeroDriveX Data Intelligence Platform

Real-data-first processing, verified provenance, and controlled synthetic augmentation. Powered by Meta Muse models. Training export is blocked unless the final dataset is ≥70% verified authentic and ≤30% synthetic.

30% cap rule: synthetic ≤ floor(3 × authentic / 7). Example: 10,000 authentic → max 4,285 synthetic. Unknown provenance never counts as authentic.

Mode: …Not signed in — datasets shown are the demo file store, not private data.

1. Create dataset

Start with a named training dataset, then select it for the steps below.

Dataset name

Dataset

2. Ingest (CSV, verified authentic)

Sample below is a demo sample, not real customer data. In production, mark AUTHENTIC_VERIFIED only when you have source + licence evidence.

3. Analyse, augment, quality, export

Each action runs against the selected dataset and reports its result below.

Synthetic rows

4. Jobs — status, progress, failures, retry

Durable jobs run on the server worker (service-authenticated; the browser never holds the worker secret). Pause, resume, and cancel are tenant-scoped. Resume also retries a failed job.

No jobs yet.

Model split — read before you send private data

muse-spark-1.3-contributor — public/demo only. Meta may use it for training. $0.10 / $0.002 cached / $0.20 per million tokens.
muse-spark-1.3 — private only. Not used for training. $1.25 / $0.15 cached / $4.25 per million tokens.
muse-image-1.0 — $0.01/image, on demand. sam-3.1 — segmentation, on demand. muse-voice-transcribe-1.0 — $0.18/hour. Multimodal ingestion is NOT operational yet — no verified endpoint functions exist, and the /api/multimodal endpoint returns 503 rather than a simulated result.
No Wonder by default — $4/$20 would burn the $150 grant.

Full split, EULA, and documentation are in the repo docs. Never paste an API key into this UI — keys are server-side only.