From store data and market evidence to a verified next step.
An AI co-worker for Indonesian sellers on Fisell, the reseller-commerce platform. It reads the store’s real numbers, researches the market live, reasons with NVIDIA Nemotron on Nebius Token Factory, and drafts one change the seller approves — then proves it was saved.
One click, no account: both stores and their 60 days of orders are synthetic. The AI calls are live once the server is configured: NVIDIA Nemotron via Nebius Token Factory.
This deployment has no NEBIUS_API_KEY yet, so analyses cannot run.
How the agent works
A bounded, multi-step workflow. The model decides what to research and what to recommend; code computes every number, runs every tool and owns every permission.
- Step 1
Read the store
SQL over the store’s own orders: cash-basis revenue, gross profit with cost frozen at payment, stock cover, checkout completion.
- Step 2
Find signals
Deterministic rules flag what moved: declining products, stock running out, thin listings, too little data to judge.
- Step 3
Plan research
NVIDIA Nemotron 3 Nano on Nebius Token Factory picks the web searches that matter for the top signal.
- Step 4
Search the market
Tavily runs those searches live. Every source keeps its URL, date and query, and is treated as untrusted text.
- Step 5
Reason over evidence
NVIDIA Nemotron 3 Ultra ranks opportunities and drafts one listing change, citing evidence ids for every claim.
- Step 6
Approve and verify
The seller edits and approves. The draft is saved as pending review, then read back from the database to prove it.
Built to be trusted with a real store
- Never changes prices, stock, orders or payments
- Never messages customers or publishes to the storefront
- Never invents numbers: figures not found in store data are flagged
- Citations to evidence that does not exist are removed
- Web pages cannot give it instructions or new permissions
- Fails loudly instead of faking a model answer