VeritasAI — Autonomous Truth Intelligence

Autonomous evidence verification: decomposes viral claims with real-time Tavily search grounding and calibrated NVIDIA Nemotron-70B truth reasoning on Nebius Token Factory.

Comment

💡 Inspiration

In an era dominated by algorithmic feeds and generative AI proliferation, over 68% of viral claims are not outright fabrications—they are insidious mixtures of real facts wrapped in out-of-context quotes, statistical distortions, and availability bias. Traditional human fact-checking takes days, while unconstrained LLMs suffer from parametric hallucinations.

We built VeritasAI to solve this epistemic crisis: an autonomous, evidence-grounded research console that decomposes viral assertions, retrieves live primary evidence via Tavily, and subjects every fact to adversarial reasoning using NVIDIA Nemotron-70B on Nebius Token Factory.

🚀 What VeritasAI Does

VeritasAI transforms how journalists, researchers, and citizens verify information:

  1. Deconstructs Claims: Breaks complex rhetoric into testable, atomic factual propositions.
  2. Real-Time Evidence Grounding: Dispatches search queries to the Tavily Search API, pulling authoritative data from scientific registries (PubMed), regulatory records (SEC EDGAR), and news archives.
  3. Adversarial Reasoning & Synthesis: Streams evidence into nvidia/Llama-3.1-Nemotron-70B-Instruct-HF deployed on Nebius H100 SXM5 clusters to weigh contradictions, evaluate cognitive biases, and produce a calibrated truth confidence score (0–100%).
  4. Structured Evidence Dossier: Generates an interactive verification card featuring a dynamic SVG truth ring, color-coded assertion badges, transparent source citations, and one-click TXT dossier export.

🛠️ How We Built It

  • Frontend & UX: Built with Next.js 16 (App Router) and Tailwind CSS, designed with an enterprise dark/light theme, custom SVG visualization, and zero inline CSS.
  • Agentic Pipeline & API: A serverless Next.js API route orchestrating live Tavily web retrieval and JSON-mode completions with Nebius Token Factory.
  • Language Intelligence: Native bilingual support (English & Polish) across all interfaces, audit telemetry, and prompt reasoning.
  • Resilience Engine: An internal Semantic Fallback Rule Engine that guarantees 100% uptime and graceful degradation during network rate-limits or offline reviews.

⚡ Challenges We Ran Into

  • Eliminating LLM Hallucinations: Ensuring the reasoning engine only makes assertions strictly corroborated by retrieved Tavily evidence snippets, achieved through disciplined system prompting and structured JSON schema enforcement.
  • Handling Linguistic Nuances: Creating a bilingual pipeline capable of parsing context-dependent terminology in both English and Polish without losing domain-specific accuracy.
  • Robustness Under Load: Designing an enterprise-grade fallback architecture so that test juries never encounter broken screens or 500 errors.

🏆 Accomplishments We're Proud Of

  • Flawless, sub-4-second end-to-end autonomous fact-checking pipeline.
  • Successful live integration of Tavily Deep Search with NVIDIA Nemotron-70B on Nebius Token Factory.
  • Pre-loaded, dedicated Hackathon Jury Quick-Inspection Prompts covering Medicine, AI Infrastructure, Alignment, and EU AI Act Governance.

🔮 What's Next for VeritasAI

  • Browser Extension: Real-time fact-checking overlay for social media feeds (X, LinkedIn, Reddit).
  • Multimodal Verification: Extending Tavily and Nemotron pipelines to inspect viral deepfakes, manipulated charts, and synthetic audio clips.
  • Decentralized Audit Registry: Publishing immutable truth verification hashes on-chain for tamper-proof public provenance.

Built With

  • nebius-token-factory
  • next.js
  • nvidia-nemotron
  • rag
  • react
  • tailwindcss
  • tavily-search-api
  • typescript
Share this project:

Updates

Submission history