Inspiration

Pickers walk into a sale with no independent rate. Dealers already know the market. Scrapwisest is meant to put a local number in the picker's pocket: ask a rate or log a sale in their language.

What it does

Chat or speak a price question, log a sale, or send a scrap photo. Place is typeahead (language, country, city). The quote is that market's currency. There is no FX. The estimator uses trusted reports (neighborhood → city → country), then a Pakistan/PKR seed if that is the market, otherwise it says it does not know. Tavily can attach same-currency /kg listings as context. It never overwrites the quote. Photos go through a vision grader (PET vs HDPE, aluminum vs steel, or refuse a selfie). English STT can stay on-device with NVIDIA Parakeet; Indic speech can use Sarvam.

How we built it

FastAPI + React/Vite. Price math is code, not the LLM. Nemotron on Nebius Token Factory: Nano extracts messy sales, Super runs chat/tools, Ultra only judges flagged anomalies. Groq grades photos. Optional Tavily + a strict same-market filter. SQLite for reports. Isolation Forest for cheap fraud flags.

Challenges

The hard part was not inventing a global ticker. Empty reports look like "no data," not a fake sale. "Rs/kg" snippets without a country used to drop; we keep them only when the request market matches. Token Factory region split (Nano EU, Super/Ultra US) and Ultra timeouts needed explicit routing and spend guards.

What we did not build

No public hosted demo. No YouTube upload yet (local docs/demo/scrapwisest-demo.mp4 only). No Bright Data. No currency conversion. No seeds for India/US/EU. No Nemotron fine-tune. No Nebius Serverless Endpoint or Job deployed. No CNN in the default requirements.txt install. History/forecast APIs are observational extras, not a live yard feed.