4real?

Is It 4real?

$467 billion in counterfeit goods.

90% of deepfakes are malicious.

Misinformation spreads 6x faster than truth.

You're about to buy a $800 Louis Vuitton bag online. Is it real?

That viral screenshot claims the president said something outrageous. Is it real?

A perfect-looking profile wants to meet you. Are they real?

Someone sends you a video of your friend saying things they'd never say. Is it real?

You can't tell anymore. And that's the problem.

Don't Get Fooled.

In a world where counterfeits cost billions and AI deepfakes fool millions...

You deserve to know what's 4real.

The Solution

4real? is the world's first multi-modal AI authenticity engine.

One platform. Four capabilities. Infinite peace of mind.

How It Works

One scan. Intelligent routing does the rest.

You don't choose what to check—our AI does it for you.

Upload anything and our intelligent routing engine analyzes your input in milliseconds:

📸 Product detected? → Authenticity verification with multi-angle analysis and trust-weighted search
🤖 AI-generated content? → Deepfake detection via custom ResNet50 model
👤 Person identified? → Background check with fakeness scoring and red flag detection
📰 Text or claim? → Fact-checking with source credibility and cross-referencing

No buttons. No menus. No confusion.

Just upload. Our AI figures out the rest.

One platform. Four capabilities. Zero tolerance for fakes.

🛍️ Product Authentication

Upload a photo. We tell you if it's fake.

Deepfake pre-screening stops AI-generated product photos
Dynamic criteria generation searches the web for brand-specific markers
Multi-angle guided capture with step-by-step instructions
Parallel AI analysis scores stitching, logos, materials, hardware
Trust-weighted reverse image search prioritizes official sources
30 seconds. Authentic or counterfeit. Done.

🤖 Deepfake Detection

Is that image even real?

Custom ResNet50 trained on 100k+ real/fake images
Instant detection before any other processing
60% threshold flags AI-generated content immediately
Stop the fakes before they fool you.

📰 Fact-Checking

Screenshot going viral? We verify it.

Extract claims with OCR + NLP
Search the web for evidence
Rate source credibility
Cross-reference multiple sources
TRUE. FALSE. PARTIALLY TRUE. You decide.

👤 Background Checks

Who are you really talking to?

Search controversies and public records
Calculate fakeness score (0-100)
Flag red flags automatically
Verify with source citations
Know who you're dealing with.

Person Coding

Inspiration

We've all been burned by fakes. A friend lost $600 on a counterfeit Supreme hoodie. Another fell for a deepfake investment scam. We watched misinformation about our campus spread faster than the truth could catch up.

The problem isn't just money—it's trust erosion. When you can't believe what you see, society breaks down.

We asked ourselves: What if authenticity could be instant? What if one platform could verify products, detect deepfakes, fact-check claims, and research people—all in real-time?

That question became "4real?"

Product Authentication Pipeline

Scan

Step 1: Item Detection We don't use hardcoded databases. Instead, Gemini Vision analyzes your photo and identifies what you're looking at—Nike Air Jordan 1, Louis Vuitton card holder, Rolex Submariner. It extracts the product name, brand, and description.

Step 2: Reverse Image Search We search Google Lens for visually similar images across the entire web. But here's the key: not all sources are equal. We built a trust-scoring algorithm:

Official brand domains get 1.0 (perfect trust)
HTTPS + verified indicators get 0.4 base + 0.15 boost
E-commerce platforms get minimum 0.5
URLs with "replica" or "fake" get -0.3 penalty
Frequent appearance across results gets +0.15 boost

We pull reference images only from high-trust sources. Your comparison isn't against random eBay listings—it's against official Nike.com product photos.

Step 3: Dynamic Criteria Generation Here's where we get adaptive. We search the web for "how to authenticate [brand] [item]" and parse authentication guides from trusted sources. The AI generates 3-10 brand-specific criteria with photo instructions:

"Front - Logo Monogram - Capture straight-on view with even lighting"
"Back - Stitching Pattern - 45-degree angle, close-up"
"Interior - Date Code - Macro shot with clear focus"

We cache these criteria using semantic similarity so similar items get instant results.

Step 4: Guided Multi-Angle Capture You follow our step-by-step instructions and capture each angle. Logo close-up. Stitching detail. Serial number. Hardware. Material texture. We guide you through exactly what authenticators look for.

Step 5: Parallel Analysis Every criterion gets analyzed simultaneously using ThreadPoolExecutor. Each one gets its own Gemini Flash call that scores 1-5 with detailed reasoning. While that's happening, we run similarity scoring:

SIFT feature matching (25% weight) - finds matching keypoints
Color histogram comparison (35% weight) - analyzes color distribution
SSIM structural similarity (20% weight) - compares image structure
Canny edge detection (15% weight) - matches edge patterns
Shape analysis (5% weight) - verifies proportions

Step 6: The Verdict We aggregate criterion scores into a total /50, calculate overall confidence percentage, and determine: AUTHENTIC, COUNTERFEIT, or INCONCLUSIVE. You get visual evidence, detailed explanations, and confidence scores for every check.

Authentic Product

Deepfake Detection Engine

Deepfake Detect

Every authentication starts here: Is this image even real? Our custom ResNet50 model examines every pixel for AI generation artifacts before any other processing begins.

The Architecture:

Custom ResNet50 trained from scratch on 100k+ real/fake images
89% accuracy across diverse image types
Pixel-level analysis for generation artifacts, inconsistencies, and AI fingerprints
Real-time inference in under 2 seconds

The Process:

Image preprocessed and normalized for model input
Deep feature extraction through 50 convolutional layers
Probability score calculated (0-100% fake likelihood)
60% threshold triggers immediate deepfake warning

The Result: If deepfake probability hits 60%, we stop everything and warn you immediately. No point authenticating a product photo generated by Midjourney or fact-checking a screenshot created by DALL-E.

REAL or FAKE. You'll know before anything else happens.

Fact-Checking Engine

False Claim

When you upload a screenshot or document, we extract text using OCR and identify factual claims with NLP. Each claim gets independently verified:

Search the web for corroborating or contradicting evidence
Evaluate source credibility (CNN: 0.8, random blog: 0.3)
Cross-reference multiple sources to build consensus
Assign verdict: TRUE, FALSE, PARTIALLY TRUE, or UNVERIFIABLE

Every claim comes with confidence scores, source citations, and explanations. No guessing. No bias. Just facts.

Background Checks

Background check

Give us a name and we search for controversies, criminal records, legal issues, and public incidents. We calculate a "fakeness score" (0-100) based on:

Presence of verified negative information
Severity of red flags (fraud, identity theft, scams)
Source reliability of findings
Recency and frequency of incidents

High-severity red flags get automatically flagged. Everything comes with source verification so you can investigate further.

The Tech Stack:

Google Gemini 2.5 Pro + Flash for vision analysis and reasoning
PyTorch ResNet50 trained on 100k+ real/fake images
Groq LLaMA 3.1 for high-speed text processing
OpenCV + scikit-image for multi-metric computer vision
SerpAPI for Google Lens reverse image search
Supabase for secure image storage
SQLite for intelligent criteria caching

The Vision

We're building a future where authenticity is instant.

Where you never wonder if it's real.

Where trust is restored, not eroded.

Where fakes don't stand a chance.

In a world full of fakes, authenticity is everything.

Real-time AI. Real information. Real confidence.

Team work

The Team Behind 4real?

We built this in 36 hours fueled by coffee, determination, and a shared frustration with getting scammed online.

From architecture planning to deployment, we divided and conquered.

Built by Youwei Zhen, Brandon Sun, Yuxin Zeng, and Krishna Mansinghka at HackHarvard 2025.

Four developers. One vision. Zero tolerance for fakes.

Don't Get Fooled. Know What's 4real.

Built at HackHarvard 2025

Try it now: 4reall.netlify.app

How we built it

Tech Stack

Frontend: Nuxt 3 + Vue 3 with real-time camera integration and responsive multi-step workflows

Backend: Flask API orchestrating multiple AI models and web search APIs

AI Models:

Google Gemini 2.5 Pro for reasoning and web search integration
Google Gemini Flash for rapid multi-criterion vision analysis
Custom PyTorch ResNet50 trained from scratch on 100k+ real/fake images (89% accuracy)
Groq LLaMA 3.1 for high-speed text processing

Computer Vision: OpenCV + scikit-image for SIFT matching, color histograms, SSIM, edge detection

APIs: SerpAPI (Google Lens), Google Search, Supabase (storage)

The Secret Sauce: Trust-scoring algorithm that prioritizes official brand sources (1.0) over sketchy resale sites (0.3). Dynamic criteria generation that searches the web in real-time instead of using hardcoded databases.

Challenges we ran into

Training a deepfake detector from scratch in 36 hours. We had to curate 100k+ images, architect the ResNet50 model, train it, and achieve 89% accuracy—all while building the rest of the platform.

Building trust scoring for unknown domains. How do you teach AI to distinguish nike.com from nike-replica-store.com? We engineered a weighted algorithm with HTTPS checks, keyword penalties, platform verification, and frequency analysis.

Real-time criteria generation. Every product is different. We couldn't hardcode authentication rules—we had to teach the system to search the web, parse authentication guides, and generate custom checklists on the fly.

Parallel processing at scale. Running 10 AI vision analyses simultaneously while keeping response times under 30 seconds required ThreadPoolExecutor optimization and intelligent caching.

Balancing speed and accuracy. We needed instant results without sacrificing verification quality. Solution: multi-metric fusion (SIFT + color + SSIM + edges + shape) weighted by reliability.

What we learned

AI alone isn't enough. The best results came from combining AI vision models, computer vision algorithms, web search intelligence, and trust scoring systems.

Context is everything. A 0.85 similarity score means something different for luxury goods versus mass-produced items. We learned to weight and interpret scores based on product type.

Speed matters. If authentication takes 5 minutes, no one will use it. We mastered parallel processing, intelligent caching, and prioritizing critical checks.

Trust is multi-dimensional. You can't just compare images—you need to evaluate source reliability, cross-reference evidence, and account for sophisticated counterfeit techniques.

Training models is hard. Getting to 89% accuracy required careful dataset curation, architecture tuning, and optimization. Every percentage point was earned.

What's next for 4real?

Blockchain authentication certificates. Mobile app with offline mode. Category expansion to pharmaceuticals and documents. Seller reputation tracking across platforms. Real-time price verification. API for marketplaces like eBay and Poshmark. Advanced deepfake detection trained on latest generative models. Community-verified authenticity database. Multi-language fact-checking.

The goal: Make authenticity instant, accessible, and irrefutable for everyone, everywhere.