Recall Radar

Inspiration

Recalled baby products keep getting resold on Facebook Marketplace, at yard sales, and in thrift stores. Recalls cover specific model numbers, but used listings almost never include them, so buyers have no easy way to check before they pay.

What it does

Paste a secondhand listing and Recall Radar finds matching CPSC recall notices, shows the evidence with exact quotes, and writes questions to send the seller, like asking for a photo of a specific label. The questions never reveal the recalled model number, so the seller reports what the label actually says. When the seller replies, the result updates. Results describe recall evidence only, never general product safety.

How we built it

NVIDIA Nemotron 3 Super on Nebius Token Factory reads listings, extracts each recall notice's affected models, dates, and identification instructions with supporting quotes, and writes seller questions. Plain Python verifies every quote against the source and makes the final match decision, with unit tests. Tavily adds manufacturer and news pages as context only. The app runs on Streamlit Community Cloud.

Challenges we ran into

The CPSC API sometimes returned its own errors disguised as recall records, and sometimes ignored search filters. Recall Radar treats those as outages, never as "no recalls," and searches a dated saved copy of 2,307 CPSC recalls so the demo stays reliable.

What we learned

AI is most trustworthy when it is boxed into clear jobs. Nemotron reads and extracts, and plain, tested code makes the decisions.

What's next for Recall Radar

Photo input for product labels (the Nemotron vision model already read a real label correctly in testing), more recall sources, and a formal evaluation.

Built With

• cpsc
• nebius-token-factory
• nvidia-nemotron
• python
• streamlit
• tavily