VeriSight AI

VeriSight AI is a multimodal verification engine that scans text, images, and documents to instantly detect deepfakes, AI-manipulated media, and viral misinformation with high-precision trust scoring.

Comment

Inspiration

In an era dominated by synthetic media, viral clickbait, and cheap-fake manipulations, distinguishing truth from fabrication has become increasingly difficult. Unverified claims, fabricated credentials on resumes, and AI-generated deepfakes spread rapidly across social platforms and hiring channels. We were inspired to build VeriSight AI—a unified, high-precision scanning engine that gives users and developers instant, transparent truth verification for any text, image, or document.

What it does

VeriSight AI is a multimodal verification platform that evaluates the authenticity and factual integrity of incoming content in real time:

  • Visual & Deepfake Analysis: Detects visual artifacts, unnatural lighting, and synthetic generation signals in uploaded images.
  • Text & Claim Verification: Audits news posts, website text, viral claims, and uploaded PDF resumes for logical contradictions, timeline mismatches, and known misinformation patterns.
  • Structured Intelligence: Delivers actionable results featuring a clear verdict (Authentic, Suspicious, or Fake), a 0–100 confidence trustScore, and a granular list of detected manipulation details.

How we built it

  • Backend Framework: Built with Node.js, Express, and TypeScript for strict type safety and scalable routing.
  • AI Engine & Models: Integrated the Groq SDK utilizing high-speed vision-language models (qwen/qwen3.6-27b) for forensic visual evaluation and text analysis.
  • Media Pipelines: Configured Multer for memory-buffered file uploads combined with Cloudinary for instant CDN hosting, enabling seamless global image access during AI inspection.
  • Robust JSON Extraction: Implemented custom balanced-brace parsing algorithms to isolate structured JSON payloads directly from model execution streams.

Challenges we ran into

  • Handling Model Reasoning Overhead: Vision-language models often append internal thinking blocks (<think>), which exceeded token limits and broke traditional JSON.parse() methods. We resolved this by building a custom character-balancing parser that isolates valid JSON objects regardless of trailing reasoning text.
  • Local to Cloud Image Passing: External AI APIs cannot access local localhost static files. Shifting from local disk storage to streaming file buffers directly to Cloudinary solved network reachability and eliminated disk I/O overhead.

Accomplishments that we're proud of

  • Zero-Latency Payload Transfer: Passing Cloudinary CDN URLs to Groq instead of heavy Base64 strings significantly reduced API payload sizes and response times.
  • Universal Input Handling: Designed a flexible single-endpoint API capable of parsing plain text, raw news snippets, image uploads, and multi-page documents seamlessly.
  • Fail-Safe Reliability: Engineered strict fallback mechanisms and schema validators that guarantee structured responses even when inputs are ambiguous or messy.

What we learned

  • Prompt Engineering Constraints: Forcing strict JSON output from reasoning models requires explicit prompt boundary rules alongside client-side string extraction logic.
  • Stream-Based Buffer Uploads: Handling file processing entirely in memory with streams creates cleaner, faster server architectures compared to temporary disk writes.

What's next for VeriSight AI

  • Browser Extension: Developing a real-time Chrome extension to highlight and fact-check text or images directly on social media feeds.
  • Web Search Grounding: Integrating live web search capabilities to cross-reference news claims against real-time global news databases.
  • Video Deepfake Detection: Expanding the vision pipeline to sample and analyze frame sequences from short-form video content.
Share this project:

Updates

Submission history