Nemotron Brief Studio

Paste a rough project brief, get a structured plan: goals, scope, milestones with size estimates, risks with mitigations, and open questions. Every section is labelled with the model that produced it.

Built on NVIDIA Nemotron open models served by Nebius Token Factory (OpenAI-compatible endpoint). Two-stage pipeline:

1. Decomposition - a large Nemotron model turns the brief into structured JSON (goals, scope, milestones, risks, open questions).
2. Metadata - a small Nemotron model drafts the title, one-line summary, tags and a per-milestone S/M/L size estimate.

The frontend renders per-section attribution badges naming the exact model id behind each part, so a reader always knows which model produced what.

Where Token Factory is used

• api/_lib/tf_client.py talks to https://api.tokenfactory.nebius.com/v1/ (/models and /chat/completions) with a bearer API key.
• api/plan.py runs both stages over that endpoint on every request.
• Model ids are resolved against the live /models list at request time (preferred Nemotron candidates first, family heuristics second) instead of being hard-coded, because catalogue ids change.

Setup

# environment
export NEBIUS_API_KEY=nf-...        # Token Factory API key
# optional overrides
export NEBIUS_BASE_URL=https://api.tokenfactory.nebius.com/v1/
export NEBIUS_REASONING_MODEL=nvidia/nemotron-3-super-120b-a12b
export NEBIUS_FAST_MODEL=nvidia/llama-3.1-nemotron-nano-8b-v1

# run the tests (stdlib only, no network, no key needed)
python3 -m unittest discover -s tests -v

# deploy (Vercel Python runtime - single entrypoint, see pyproject.toml)
npx vercel deploy --prod --yes

No third-party Python packages are required: the API and tests are stdlib-only Python 3.9+. The Vercel Python runtime builds the project as a single Python app with the entrypoint declared in pyproject.toml (api.app:handler routes /api/health and /api/plan; static files are served by the CDN).

Sample mode

If the service is not connected (no API key configured, upstream auth, rate limit, upstream outage or unparsable model output), the API returns an HTTP 503 with a machine-readable error code. The frontend then renders a clearly labelled illustrative sample so the interface shape stays visible - it never presents sample content as model output, and genuine client input errors (HTTP 400/413) never trigger the sample path.

API

• POST /api/plan - body {"brief": "...", "depth": "standard"|"deep"}. Returns {"mode": "live", "result": {...}} where result.attribution names the model behind each stage.
• GET /api/health - liveness plus whether the key is configured.

Limits

• Briefs up to 20,000 characters.
• The metadata stage degrades gracefully: if the small model fails, the plan still ships with brief-derived defaults and the attribution states it.
• One repair retry if the reasoning stage returns unparsable JSON.

License

MIT - see LICENSE.

Built With

• html
• javascript
• nebius-token-factory
• nvidia-nemotron
• python
• vercel