Nebius and Nemotron
How wovain uses NVIDIA Nemotron on Nebius Token Factory to write briefs, judge designs, write ad copy and decide in autopilot.
Wovain calls NVIDIA Nemotron 3 Super (nvidia/nemotron-3-super-120b-a12b, the NEMOTRON_MODEL constant in packages/core/src/nemotron.ts) on Nebius Token Factory. The endpoint is OpenAI-compatible: NEBIUS_BASE_URL, https://api.tokenfactory.nebius.com/v1/ by default, authenticated with NEBIUS_API_KEY. Every reply is validated with a Zod schema.
Nemotron does four jobs. It never replaces the checks that code makes.
1. Design brief and listing
draftBrief turns an idea into the art prompt, title, description and garment colors. Names Qloo matched for the seed are passed as names never to print.
2. Design judge
judgeDesign receives the brief, listing and artwork facts and returns a verdict (ready or revise), a 0 to 100 score, a pass or fail check for title, description, idea fit and originality, up to five fixes and a one-line coach note.
The print-size rule is decided by code (printFacts, applyFacts): art narrower than 2880 px fails whatever the model says, and the verdict becomes revise.
3. Ad copy
writeAdCopy writes up to three ad variants with an angle, a headline and a message for the Meta campaign.
4. Autopilot
With autopilot on, pickIdea chooses the most promising ranked idea and reviewDesign approves or rejects the design, each with a one-line reason. Code still gates the run:
| Gate | Enforced by code |
|---|---|
| Judge | A design the judge says to revise is never approved, whatever Nemotron says |
| Idea choice | An index the list does not have falls back to the top-ranked idea |
| Price | At least $10 profit and a 30% margin after card fees, sales tax and a refund reserve (priceFor) |
| Daily budget | Between $1 and $500 |
| Meta | Every object is created paused |
The reason for each decision is stored on the design, and a run records its step and any error.
Why Nemotron
- A second model with a fixed rubric and a schema catches what a host agent's self-review misses.
- Structured output at a low cost per call suits several small decisions per run.
- Running the decisions on Token Factory means they do not depend on a host agent staying in the loop.
Configuration
NEBIUS_API_KEY=...Without the key, steps that need Nemotron fail with a NOT_CONFIGURED error naming the variable. Nothing else breaks.