Bid Scout

• 

• This image is currently processing. Please be patient...
  

Inspiration

We are Design Master Solutions, a small web and app agency in Massachusetts. Public work (town websites, agency portals) is out there, but it is spread across thousands of procurement pages, most search hits are closed or off-trade, and when we tried letting a model skim them it happily invented a fit. We wanted a short list we can trust, not a long list we have to re-check.

What it does

Give Bid Scout a company profile: services, region, the facts it may claim, and what the company cannot do. It

1. searches the live web with Tavily for open RFPs and RFQs that match each service and region;
2. reads each solicitation in full with Tavily Extract;
3. asks NVIDIA Nemotron 3 Ultra, served by Nebius Token Factory, to judge it: is it an open solicitation, when is it due, how well does the company fit (0-100), what are the gaps, and which sentences on the page prove each claim;
4. checks every quoted sentence against the page text. Quotes that are not on the page are dropped, and a fit with no verified quote is capped at 30;
5. drops closed bids, ranks the rest, and drafts a response for the best one using only the company's facts. Anything the facts do not cover becomes a [CONFIRM: ...] marker plus a go/no-go checklist.

Real run on Token Factory (29 Oct 2026): 3 searches, 6 solicitations read, 4 dismissed, 2 ranked (a town website RFP at 72 and a Massachusetts services RFP at 55). 19 model quotes were not on the page and were thrown away; one website RFP had 8 quotes and none were real, so it was capped at 30.

How we built it

Node 22, no dependencies. src/scout.mjs is the agent, src/clients.mjs talks to Tavily and to the OpenAI-compatible Token Factory endpoint (nvidia/Nemotron-3-Ultra-550b-a55b), src/runner.mjs records and replays every external call so judges can run npm run demo without keys. Six offline tests cover the filters, the quote check and the fit threshold; both safety checks were seen failing before they passed.

Challenges

Models quote confidently. Making the quote check strict (case and whitespace are normalized, but the words must be on the page) and tying the fit score to it was the core of the project. Switching providers we also hit a case-sensitive model id (model_not_found) and fixed it by listing /v1/models.

Accomplishments

A ranked list where every "why it fits" line has evidence you can find on the bid page, and a draft that never invents references, pricing or insurance.

What we learned

Grounding beats prompting: a verifiable rule (quote must exist) did more for trust than any instruction in the prompt.

What's next

Run it daily for our own pipeline, add state procurement portals as sources, and email the morning short list.

Feedback on Nebius Token Factory and NVIDIA

• The OpenAI-compatible endpoint was a true drop-in: the same client runs on build.nvidia.com and Token Factory, only base URL, key and model id change.
• Model ids are case-sensitive and differ between NVIDIA's API (nvidia/nemotron-3-ultra-550b-a55b) and Token Factory (nvidia/Nemotron-3-Ultra-550b-a55b). Accepting ids case-insensitively, or listing the NVIDIA alias, would save a round trip for every team that switches providers.
• Nemotron 3 Ultra returns reasoning in reasoning_content and the answer in content, which kept JSON parsing simple.

Built during the Submission Period

Whole project; first commit 28 Oct 2026.

Built With

• javascript
• nebius-token-factory
• node.js
• nvidia-nemotron
• tavily