Inspiration
Invoice fraud works because documents look real. A look-alike sender domain or a quietly changed bank account is easy to miss by hand, and asking an AI "is this legit?" just moves the trust problem to the model. We wanted a verifier whose verdicts can be checked by a human.
What it does
VerifyFlow takes a document such as an invoice, extracts every checkable claim (vendor, bank account, email domain, amounts), verifies each claim against live evidence, and produces an auditable trust report.
- Evidence decides, not the model. The model proposes findings, but deterministic code makes every verdict. Proposals the evidence can't back up are rejected, and the rejection is shown in the report.
- Deterministic checks catch things like look-alike domains that use characters from other alphabets.
- Per-claim proof: each flag links to a quote, a source or a check you can audit.
- Challenge Mode: five adversarial documents show the verifier refusing to be fooled.
How we built it
- Nebius Token Factory runs all inference.
- NVIDIA Nemotron models are used in tiers: a fast model extracts claims, and stronger models handle harder judgments.
- Tavily provides live search for evidence.
- Every result shows the model used, tokens and latency, at about [X] cents per document.
- A regression suite runs the five challenge documents so changes can't silently break a verdict.
- Deployed on Vercel, with a Docker build for self-hosting.
Challenges we ran into
- Keeping the model useful without letting it overrule evidence.
- Making slow, multi-step runs (15-60 seconds) feel clear, so we added saved example results and a progress view.
- Staying honest about limits: on our own challenge set a plain LLM review also flagged the problems, so our edge is auditability, not just detection.
Accomplishments we're proud of
- Rejected model proposals are visible to the user.
- Every verdict traces back to evidence.
- The whole flow runs on Nebius and Nemotron with transparent cost.
What we learned
Cheap models do the wide extraction work well, and spending a stronger model only on ambiguous judgments keeps cost low. Showing why something was rejected builds more trust than showing a confident answer.
What's next
More document types, more deterministic checks, saved audit trails per vendor, and integrations with accounts-payable tools.
Hackathon disclosure
[Pick one: "VerifyFlow was built entirely during the hackathon period." OR "The project started on [date]. During the hackathon we added Nebius Token Factory and Nemotron tier routing, Tavily evidence checks, Challenge Mode, the web UI and the telemetry view."]
Try it
Built With
- fastapi
- html
- javascript
- nebius-token-factory
- nvidia-nemotron
- python
- react
- tavily
- vercel