FactCheck Radar

Inspiration

In an era of hyper-fast information velocity, synthetic media, and viral misinformation, verifying claims in real time remains a critical engineering and societal bottleneck. Existing fact-checking workflows rely heavily on manual verification, lagging behind the speed at which misinformation propagates across social channels. We set out to engineer FactCheck Radar: an enterprise-grade, autonomous multi-agent intelligence platform designed to deliver deterministic, evidence-backed fact-checking with transparent reasoning and calibrated confidence scoring.

What It Does

FactCheck Radar orchestrates a specialized multi-stage agent pipeline that autonomously analyzes claims:

1. Extraction & Normalization: Deconstructs raw claims or articles into atomic verifiable assertions.
2. Query Generation: Formulates precision search queries optimized for search engines.
3. Web Evidence Retrieval: Executes non-blocking parallel searches via Tavily AI with domain deduplication and SSRF hardening.
4. Stance Assessment: Evaluates evidence against assertions (supports, contradicts, neutral).
5. Verdict Synthesis & Confidence Calibration: Synthesizes final verdict, reasoning breakdown, and calibrated confidence score.

The frontend features a real-time Server-Sent Events (SSE) streaming timeline, interactive confidence meters, expandable source citations, preset claim triggers, and a polished dark/light mode toggle built with React and Tailwind CSS v4.

How We Built It

• Backend: FastAPI (Python 3.11) utilizing asynchronous architecture (asyncio.to_thread), LRU TTL caching, and robust SSRF / path-traversal protection.
• AI & LLM Integration: Nebius Token Factory OpenAI SDK wrapper targeting https://api.tokenfactory.nebius.com/v1 with startup /v1/models discovery, automatic fallback chains, and self-healing JSON retry logic.
• Model Synergy: Dual-model architecture leveraging NVIDIA Nemotron Nano 30B (nvidia/Nemotron-3-Nano-30B-A3B) for fast extraction and query generation, and NVIDIA Nemotron Ultra 550B (nvidia/Nemotron-3-Ultra-550b-a55b) for deep synthesis and reasoning.
• Search Engine: Tavily AI API for real-time web evidence collection.
• Frontend: React, Vite, Tailwind CSS v4, Lucide icons, and SSE streaming client.
• Testing & DevOps: 13/13 passing Pytest unit tests covering schemas, cache, pipeline flow, and SSRF filters; multi-stage Docker container (node:20-alpine + python:3.11-slim) and docker-compose.yml for 1-command deployment.

Challenges We Ran Into

• API Latency vs. Reasoning Depth: Balancing rapid token parsing with heavy synthesis required a hybrid model strategy—delegating extraction/query generation to Nemotron Nano 30B while reserving Nemotron Ultra 550B for complex synthesis.
• JSON Parsing Robustness: LLM output variance in multi-agent pipelines necessitated self-healing JSON repair logic with in-band prompt correction and strict Pydantic schema validation.
• Real-Time Streaming: Implementing reliable Server-Sent Events over FastAPI to stream live agent progress without blocking the async event loop.

Accomplishments That We're Proud Of

• Built a fully autonomous, production-grade 5-stage agent pipeline running on Nebius Token Factory and NVIDIA Nemotron.
• Achieved 100% test coverage on core backend modules with 13/13 passing Pytest tests.
• Delivered a polished, responsive streaming UI with real-time SSE progress indicators and expandable source cards.
• Engineered enterprise-grade security hardening including SSRF protection against loopback/internal IPs and path traversal defense.

What We Learned

• The power of model specialization: pairing a fast 30B model with a reasoning 550B model creates an optimal balance of throughput and depth.
• Effective prompt engineering and robust fallback handling are paramount for deterministic multi-agent orchestration.

Built With

• docker
• fastapi
• nebius-token-factory
• nvidia-nemotron
• python
• react
• tailwind-css
• tavily