EvidencePilot

• 

• 

• 

• 

Evidence before certainty , AI-powered claim investigations with traceable sources and human led review.

Category: Best Apps and Agents

Inspiration

Misinformation spreads quickly, and AI-generated answers can sound convincing even when the evidence is incomplete or misleading.

We wanted to build a platform that helps people investigate factual claims before sharing or publishing them. Instead of presenting a confident answer without sufficient context, EvidencePilot makes sources, limitations, and uncertainty visible.

Our goal is to make evidence based research more accessible while keeping humans responsible for final decisions.

What it does

EvidencePilot is an AI-powered research and fact-checking assistant.

Users can paste a statement or article and choose between Quick Check and Deep Investigation.

The platform:

• Extracts up to three independently checkable factual claims.
• Searches the web for relevant evidence using Tavily.
• Uses NVIDIA Nemotron to assess each claim against retrieved sources.
• Classifies findings as supported, contradicted, mixed, or insufficient evidence.
• Displays citations, source relationships, limitations, and investigation details.
• Suggests cautious wording when a complete rewrite is supported.
• Provides a human editorial workspace for reviewing sources and editing drafts.
• Supports exporting, reopening, and comparing evidence reports.

EvidencePilot does not claim that AI-generated verdicts are independently proven facts. Users are encouraged to inspect original sources.

How we built it

We developed EvidencePilot using Next.js 15, React 19, and TypeScript, with server-side AI orchestration.

The AI workflow runs NVIDIA Nemotron 3 Super 120B A12B through Nebius Token Factory.

Nemotron handles claim extraction, research planning in Deep Investigation, evidence assessment, and cautious rewriting.

Tavily Search retrieves live web information. Our application maps model-selected citation indices back to URLs actually returned by the search provider rather than accepting invented model-generated links.

We also implemented source validation, conservative uncertainty handling, bounded provider calls, rate limiting using Upstash Redis, and a responsive research workspace.

The application is hosted on Vercel and the source code is publicly available under the MIT license.

Challenges we faced

One major challenge was ensuring that confident AI output did not misrepresent incomplete evidence.

We addressed this through deterministic citation validation, source-relationship checks, and safeguards that downgrade unsupported conclusions to insufficient evidence.

Other challenges included handling provider failures, limiting expensive API requests, preserving research during imports, and preventing long documents from being silently truncated during rewriting.

We also separated machine-generated assessments from human editorial decisions to maintain transparency.

Accomplishments we're proud of

We built and deployed a working platform rather than a prompt only prototype.

EvidencePilot includes two distinct AI investigation workflows, live web retrieval, source linked findings, an editorial review system, structured report exports, and safeguards against unsupported certainty.

The project also has automated testing, continuous integration, and a publicly accessible demonstration.

What we learned

We learned that building a responsible AI research tool requires more than connecting a language model to a search API.

Source provenance, uncertainty communication, reliable failure handling, and human oversight are essential.

We also learned that an AI system's ability to produce convincing explanations is different from its ability to establish factual truth.

What's next

We want to expand EvidencePilot with broader multilingual evaluation, stronger time sensitive research, better source quality assessment, and a more comprehensive factual evaluation dataset.

Longer term, our vision is to build an evidence intelligence platform that helps writers, researchers, educators, and organizations make more informed decisions.

Built With

• chatgpt
• codex
• nebius
• next
• nvidia
• react
• typescript
• vercel