FlowPilot Verify

Inspiration

AI-generated research can sound convincing while containing unsupported claims. We wanted to build a research copilot that does more than generate answers—it checks whether its claims can be traced to actual web sources.

What it does

FlowPilot Verify turns a research question into a synthesized answer with claim-level evidence. It retrieves web sources using Tavily, extracts claims and supporting quotations, and checks whether each quotation genuinely appears in the retrieved source content. NVIDIA Nemotron then evaluates whether the evidence semantically supports each claim.

The interface presents individual claims, provenance checks, semantic verdicts, evidence quotations, source links, and research limitations in one place.

How we built it

We built FlowPilot Verify using Next.js, React, TypeScript, Tailwind CSS, and Zod. The backend integrates NVIDIA Nemotron through Nebius Token Factory for AI reasoning and Tavily for web research.

Our central design principle is simple: code verifies provenance; Nemotron evaluates meaning. Deterministic checks gate semantic evaluation, preventing unsupported quotations from being treated as verified evidence.

Challenges we faced

One challenge was distinguishing a quotation that genuinely appears in a retrieved source from one that merely sounds plausible. We implemented Unicode, punctuation, and whitespace normalization to make quotation matching more robust. We also added schema validation and automated tests to check provenance gating and structured model outputs.

What we learned

We learned that source traceability and semantic correctness are different problems. A quotation can be authentic without fully supporting a claim, and an AI verdict is not a guarantee of truth. FlowPilot Verify makes these distinctions visible so researchers can inspect evidence and understand limitations rather than blindly trust generated answers.

What's next

We aim to improve source coverage, add broader evaluation datasets, refine contradiction detection, and measure verification performance against manually labelled claims.