TrueSight AI

See what’s real. Trust what’s true.

Comment

Inspiration

In today’s world, AI-generated images, deepfake videos, and synthetic media are everywhere — on Instagram, YouTube, WhatsApp, and even news outlets. Most people can no longer tell what is real and what is AI-made. This creates serious risks: misinformation spreading faster than truth fake political videos influencing public opinion deepfake job scams & fraud manipulated evidence in personal disputes AI celebrity voice scams trust erosion across society I wanted to build a tool that gives ordinary people the power to verify digital content instantly. That’s how TrueSight AI was born — a platform that helps anyone see the truth behind any image or video.

What it does

TrueSight AI analyzes media and detects whether it is real or AI-generated, using multiple AI models and forensic techniques.

Key capabilities: Detects AI-generated images Identifies deepfake faces in videos Highlights manipulated or forged regions Extracts and analyzes metadata (EXIF, compression artifacts) Predicts the possible AI model used (Midjourney, Sora, etc.) Generates a Truth Score from 0 to 100 Accepts uploads or URLs (YouTube, Instagram, websites) Users get a clean, simple dashboard to understand the authenticity of the content in seconds.

How we built it

Frontend (React + Tailwind + shadcn/ui) We designed a fast, modern UI with: drag-and-drop upload live analysis results heatmap visualization responsive, minimal design reusable UI components

Backend (Node.js) The backend handles: secure file uploads media URL extraction communication with the Python AI engine standardized result formatting caching and job control

AI Engine (Python + FastAPI) This is the core intelligence of TrueSight AI. We built an ensemble pipeline combining: CLIP anomaly detection GAN fingerprint analysis Deepfake face landmark detection Image noise-level inconsistency scanning Forgery & splice detection models Metadata scanning Video keyframe extraction We aggregate model results into a unified probability score The output is converted into a user-friendly Truth Score.

Challenges we ran into

  1. Video processing was slow Deepfake detection on multiple frames required optimization. We solved it by sampling fewer frames intelligently and parallelizing inference.

  2. False positives on artistic images Paintings and illustrations confused models. We tuned thresholds and added metadata + texture analysis.

  3. Handling social media URLs Instagram and YouTube constantly change their HTML structure. We built fallback scrapers and validation layers.

  4. Combining multiple model outputs Creating a meaningful “Truth Score” was tricky. We created a weighted ensemble model to merge multiple detection signals.

Accomplishments that we're proud of

Built a fully working AI real-vs-fake detection engine Designed a clean, modern, user-friendly UI Achieved accurate detection on viral AI content Successfully extracted deepfake heatmaps for faces Created a robust pipeline that works on images and videos Developed a simple representation (Truth Score) from complex AI analysis Built a system that could genuinely help people avoid misinformation

What we learned

How deepfakes are created and how to detect them Techniques for analyzing GAN fingerprints and noise patterns Efficient video frame extraction and processing Multi-service architecture with React → Node → Python Importance of good UI/UX for trust and safety tools How to translate complex AI analysis into simple explanations for everyday users That truth verification is becoming an essential part of digital life

What's next for TrueSight AI

  1. Browser Extension Instantly verify images/videos on Instagram, YouTube, Facebook, and Twitter.

  2. Real-time scanning on live video calls Protect users from deepfake Zoom/WhatsApp scams.

  3. Public API for journalists & fact-checkers Allow third-party platforms to integrate TrueSight AI.

  4. Mobile App (Android + iOS) Tap → Scan → Verify.

  5. Blockchain-backed authenticity certificates Store verified content fingerprints on-chain.

  6. Advanced source model attribution Tell users exactly which AI model generated the content.

  7. Misinformation monitoring dashboard Track trends of AI-generated fake media globally.

Built With

Share this project:

Updates