About

Built for the offer that expires tonight.

Most job and client scams can be caught with public evidence. Nobody has time to find it at 11pm, so ProofBeforePay sends seven agents to find it for you, and shows you every piece.

The recruiter writes from a domain registered last week. The company's own careers page says it never asks candidates for money. The same message is already on a scam forum. The evidence is out there. It just takes an hour to find, and the offer expires tomorrow. So we built agents that do the digging in under a minute, and show every piece of their work.

$0M

lost to job scams in the US in 2024 (FTC)

0

steps in every check, from reading the message to arguing with the verdict

0

NVIDIA Nemotron models, each given the work that fits its size

0¢

in model costs for a full check, measured on every run

0%

cheaper than sending everything to the largest model

[01]Who it's for

For anyone with a message that feels a bit off.

Job seekers

Remote roles that pay too well, recruiters who only talk on Telegram, onboarding fees that are "refundable". Check before you reply.

Freelancers

Clients who overpay by cheque and ask for the difference back, or want you to buy software from their vendor. Check before you start work.

Students and first jobs

The people targeted most, and the least likely to know what a real hiring process looks like. Every report explains its reasoning.

[02]Honest by design

Built to be careful, not just confident.

A wrong “this is a scam” can cost someone a real job. A wrong “looks fine” can cost them their savings. These rules are enforced in code.

  1. 01

    Missing evidence isn't proof of fraud.

    If the web has nothing on a company or a person, the claim is marked inconclusive. It is never counted as contradicted.

  2. 02

    Quotes are checked against the page.

    A quote only appears in the report if those exact words are on the source it cites.

  3. 03

    Money requests can't get a green light.

    If the message asks for money, a rule in code blocks a Proceed verdict, whatever the model says.

  4. 04

    Signals, not accusations.

    It never calls a company or a person a scammer. When a brand is being impersonated, the real company isn't the one at fault.

[03]Under the hood

Built on tools that show their work too.

Every model call is metered, every web call is counted, and every report shows both. Nothing here is a black box.

NVIDIA Nemotron 3

Nano reads and checks, Super re-checks and challenges, Ultra weighs everything once.

Nebius Token Factory

Runs all three Nemotron models behind one OpenAI-compatible API, metered per call.

Tavily

Search, exact-phrase search, careers-site crawling, and reading pages so we never open them ourselves.

Supabase

Postgres for accounts, saved investigations, the evidence trail and the spending ledger.

Resend

Delivers the one-time codes that verify your email and reset your password.

Next.js and FastAPI

One container: the site you're reading, and the API that runs the agents and streams them live.

Open source, MIT licensed

Built during the Nebius × NVIDIA Global AI Hackathon, October 2026. Read the code, run it yourself, or improve it.

Got a message that feels a bit off?

Check it before you reply, send your documents, or pay for anything. It's free, and it shows its work.