Evidence Verification Agent
An AI-powered multimodal evidence verification platform for assessing the authenticity of news and video content, integrating deepfake detection, evidence-backed article verification, and real-time browser extension analysis.
Overview
Evidence Verification Agent is a multimodal evidence verification platform. It analyzes video and news content for authenticity signals and surfaces them to users in real time through a browser extension.
Screenshots
Problem
Deepfakes and viral misinformation are consumed on platforms where users have seconds to judge credibility. Existing verification tools are manual, slow, and out of reach for the typical reader or viewer.
Architecture
Solution
Evidence Verification Agent runs CNN-based video deepfake detection and LLM-assisted news credibility assessment through a single pipeline. An evidence retrieval stage grounds assessments in sources, and a Chrome extension brings the results to the user's current tab.
Key Features
- Video deepfake detection using CNNs
- AI-assisted news credibility assessment
- Evidence-backed article verification
- Real-time browser extension integration
- Multimodal verification pipeline
- Source attribution for analyzed news content
Technical Decisions
- CNNs for video authenticity classification, focusing on spatial artifacts.
- LLM-assisted assessment for news credibility with an evidence-retrieval grounding stage to reduce hallucinated reasoning.
- A browser extension as the delivery surface because that is where misinformation is actually encountered.
- FastAPI backend to keep the pipeline callable from both the extension and a web dashboard.
Lessons Learned
- Verification is a trust product — confidence, provenance, and sourcing matter as much as the model output.
- Grounding LLM assessments in retrievable evidence is non-negotiable for anything presented as fact.
- Prototype first, prove the signal, then scale the infrastructure.
Future Roadmap
- Publish evaluation metrics for the CNN model
- Expanded evidence sources and better source attribution
- Support for additional video and audio modalities
- A public demo and dataset release