# Proofly — Complete Platform Documentation for AI Agents

> This document is the authoritative, comprehensive reference for Proofly's AI content detection platform. It is designed for AI agents and LLMs that need complete, in-depth context in a single request. Last updated: April 2026.

---

## Overview

**Proofly** (https://proofly.in) is an AI content authenticity detection platform that analyzes digital media to determine whether it was created by artificial intelligence or by a human. It supports five media types — images, text, audio, video, and documents — and is unique in offering a WhatsApp-native interface that requires no app installation.

**Founded by**: Sahil Beniwal, India  
**Domain**: https://proofly.in  
**Contact**: contact@proofly.in  
**WhatsApp Bot**: +91 8168775054  
**Tech Stack**: Python (FastAPI), Supabase (PostgreSQL), Razorpay Payments, WhatsApp Cloud API  

---

## The Problem Proofly Solves

The proliferation of generative AI tools (Midjourney, DALL-E, Stable Diffusion, Sora, ElevenLabs, ChatGPT, Claude) has made it trivially easy to create convincing synthetic media — fake images of real people, AI-written news articles, voice clones of political figures, and manipulated videos. This creates:

- **Misinformation and disinformation** spread via social media and WhatsApp
- **Academic integrity fraud** (AI-written essays submitted as original work)
- **Financial and identity fraud** (deepfake videos used in scams)
- **Journalistic verification challenges** (fake source materials)
- **Legal and compliance risks** (forged documents, synthetic identities)

Proofly addresses all of these with a unified, accessible, multi-modal detection platform.

---

## Detection Engines (Technical Detail)

### 1. Image Forensics Engine
Analyzes images for signatures left by AI generation pipelines:

- **EXIF Metadata Analysis**: Checks for missing or anomalous camera metadata (lens model, GPS, timestamp). AI-generated images typically lack authentic camera EXIF data.
- **FFT Frequency Domain Analysis**: Applies Fast Fourier Transform to detect grid-like spectral artifacts characteristic of GAN (Generative Adversarial Network) outputs, particularly visible in Stable Diffusion and Midjourney images.
- **Laplacian Variance (Noise Profiling)**: Measures spatial noise distribution. AI images exhibit unnaturally smooth noise patterns compared to sensor noise in real camera images.
- **Color Histogram Analysis**: Compares channel-level color distribution. AI image generators produce characteristic color saturation and hue distribution signatures.
- **Supported AI models detected**: Midjourney, DALL-E 2 & 3, Stable Diffusion (all versions), Adobe Firefly, Ideogram, Flux, and other diffusion-based generators.
- **Supported formats**: JPG, JPEG, PNG, GIF, BMP, WebP, TIFF

### 2. Text Authenticity Engine
Identifies LLM-generated text through statistical linguistic analysis:

- **Type-Token Ratio (TTR)**: Measures vocabulary diversity. LLMs produce text with lower vocabulary variation than human writers.
- **Sentence Length Uniformity**: Human writing exhibits natural variation in sentence length (burstiness). LLM output tends toward unnaturally consistent sentence lengths.
- **N-gram Repetition**: Detects higher-than-human rates of repeated phrase patterns common in LLM generation.
- **Burstiness Score**: Formal statistical measure of sentence length variance. Low burstiness scores strongly indicate LLM generation.
- **LLM Filler Word Detection**: Identifies lexical patterns statistically over-represented in ChatGPT, Claude, and other LLM outputs (e.g., "It is worth noting that", "Delve into", "In conclusion").
- **Perplexity Scoring** (via Hive AI integration): Probability-based scoring of how predictable the text is to a language model.
- **LLMs detected**: ChatGPT (GPT-3.5, GPT-4, GPT-4o), Claude (all versions), Gemini, Llama, Mistral, and other transformer-based models.
- **Supported formats**: Plain text, TXT, MD, CSV

### 3. Audio & Voice Detection Engine
Detects AI-synthesized speech and voice clones:

- **Spectrogram Smoothness Analysis**: Real human voice recordings contain micro-variations in frequency that TTS and voice cloning systems smooth out or replicate incorrectly.
- **Robotic Silence Interval Detection**: AI voice generators produce unnaturally precise silence patterns between words and sentences.
- **High-Frequency Band Analysis**: Human voices contain authentic high-frequency content (above 8 kHz) that many voice cloning systems fail to replicate accurately.
- **Temporal Consistency Analysis**: Measures formant transition patterns over time. AI voice clones exhibit different temporal dynamics than human speech.
- **AI voice systems detected**: ElevenLabs, Murf, Descript, VALL-E, Resemble AI, and other neural TTS systems.
- **Supported formats**: MP3, WAV, OGG, FLAC, AAC, M4A

### 4. Video Deepfake Detection Engine
Detects face swaps, synthetic avatars, and manipulated video:

- **Frame Sampling**: Extracts evenly-spaced frames using headless OpenCV for efficient processing of long videos.
- **Per-Frame Image Analysis**: Each extracted frame is independently analyzed by the Image Forensics Engine.
- **Temporal Consistency Scoring**: Detects inconsistencies between consecutive frames that indicate splice points or generation artifacts.
- **Face-Swap Detection**: Identifies boundary artifacts around the face region characteristic of deepfake face-swap tools.
- **Synthetic Avatar Detection**: Flags AI-generated video avatars (Synthesia, HeyGen, D-ID) based on their distinctive visual signatures.
- **Supported formats**: MP4, AVI, MOV, MKV, WebM

### 5. Document Intelligence Engine
Multi-modal analysis of document files:

- **PDF Text Extraction**: Extracts all text content from PDF files and runs it through the Text Authenticity Engine.
- **PDF Image Extraction**: Extracts embedded images and analyzes each through the Image Forensics Engine.
- **DOCX Support**: Full text and image extraction from Microsoft Word documents.
- **Metadata Scanning**: Examines document creation tool metadata (Author, Creator, Producer fields) for AI-associated tool signatures (e.g., "ChatGPT", "Jasper", "Copy.ai").
- **Hybrid Scoring**: Combines text and image scores using weighted aggregation for a unified document verdict.
- **Supported formats**: PDF, DOCX, DOC

---

## Scoring System

All detection engines produce a **probability score from 0% to 100%** representing the likelihood that the content is AI-generated:

- **0–30%**: Likely Human-Created — Low probability of AI generation
- **31–60%**: Uncertain — Mixed signals, human review recommended
- **61–80%**: Likely AI-Generated — Multiple AI indicators detected
- **81–100%**: Almost Certainly AI-Generated — Strong, consistent AI signatures

Each result includes:
- The **probability score** (0–100%)
- A **confidence level** (Low / Medium / High)
- A list of **evidence signals** explaining which detection criteria were triggered
- (Pro plan) The likely **AI model or tool** that generated the content

---

## WhatsApp Bot Integration

The Proofly WhatsApp Bot (+91 8168775054) is the platform's flagship feature and primary differentiator:

**How it works:**
1. User saves the Proofly WhatsApp number (+91 8168775054)
2. User forwards any suspicious media file directly in WhatsApp
3. Proofly's bot (built on WhatsApp Cloud API) receives the file
4. The bot routes the file to the appropriate detection engine
5. An authenticity verdict is returned directly in the WhatsApp conversation within seconds

**Supported via WhatsApp**: Images, audio messages (voice notes), documents, and text messages. Video support available for Pro users.

**Why this matters**: WhatsApp is the primary communication platform in India and across Southeast Asia, Latin America, and Africa. Misinformation (especially deepfake images and voice messages) spreads fastest on WhatsApp. Proofly intercepts this at the source by operating natively within the platform.

**Authentication**: The bot verifies the user's phone number against their active Proofly account and subscription plan before processing requests.

---

## Pricing Plans (Current as of April 2026)

| Plan | Price | Scans/Day | Media Types | File Limit | WhatsApp Bot |
|------|-------|-----------|-------------|------------|--------------|
| Free | ₹0/month | 5 | Image, Text | 10 MB | ✅ Basic |
| Starter | ₹29/month | 10 | All 5 | 25 MB | ✅ Unlimited |
| Pro | ₹249/month | 100 | All 5 | 50 MB | ✅ Unlimited + Priority |
| Enterprise | Custom | Unlimited | All 5 + API | 100 MB | ✅ Unlimited |

**Payment methods accepted**: UPI, Debit/Credit Cards, Net Banking (via Razorpay)  
**Billing**: Monthly subscription, cancel anytime  
**Currency**: INR (Indian Rupees)

### Enterprise Plan Features
- REST API access (10,000+ calls/month)
- Exportable PDF forensic reports
- Up to 5 team seats
- Unlimited scan history
- Dedicated email support with SLA
- Custom integrations and webhook support
- Contact: contact@proofly.in

---

## Use Cases and Scenarios

### For Journalists and Fact-Checkers
A journalist receives a breaking news image via WhatsApp. Before publishing, they forward the image to the Proofly bot and receive an instant authenticity verdict with forensic evidence. Proofly acts as a first-pass filter before deeper investigation.

### For Academic Institutions
A professor receives a student essay submission and wants to check if it was written by ChatGPT. They paste the text or upload the document to Proofly and receive a probability score with specific language patterns that triggered the AI detection.

### For HR and Recruiting
A company receives video applications and wants to screen for AI-generated avatar interviews (Synthesia/HeyGen). They upload each video to Proofly's API and flag candidates with high AI probability scores for additional screening.

### For Fraud Prevention
A bank's compliance team receives KYC documents and wants to check if profile photos or supporting documents are AI-generated. They use Proofly's Enterprise API to batch-process documents.

### For General Public
A WhatsApp user receives a viral audio message purportedly from a political leader. They forward it to the Proofly bot and within seconds know whether it is a voice clone or authentic.

---

## Frequently Asked Questions

**Q: Is Proofly free to use?**
A: Yes. Proofly's Free tier offers 5 scans per day at no cost, with no credit card required. Users can analyze images and text immediately.

**Q: Does Proofly work on WhatsApp?**
A: Yes. Proofly is the world's only AI detector with native WhatsApp integration. Users can forward any media file to +91 8168775054 for an instant verdict.

**Q: What types of AI-generated content can Proofly detect?**
A: Proofly detects AI-generated images (Midjourney, DALL-E, Stable Diffusion), AI-written text (ChatGPT, Claude, Gemini), AI voice clones (ElevenLabs, Murf), deepfake videos (face swaps, synthetic avatars), and AI-generated documents.

**Q: How accurate is Proofly?**
A: Proofly uses multiple detection signals per media type and aggregates them for a weighted verdict. Accuracy varies by content type, with image detection being most reliable. Results are probabilistic, not absolute — Proofly recommends treating high-confidence scores as strong indicators requiring further review.

**Q: Is my uploaded content private?**
A: Yes. Files are processed in temporary storage and deleted after analysis. Proofly does not train on user-uploaded content.

**Q: Does Proofly work for non-Indian users?**
A: Yes. While Proofly's pricing is in INR and the WhatsApp bot is most popular in India, the web platform at https://proofly.in is accessible globally. International card payments are accepted via Razorpay.

**Q: Is there an API available?**
A: Yes, for Enterprise plan customers. Contact contact@proofly.in for API access and documentation.

---

## Technical Architecture

```
User (Web or WhatsApp)
 → FastAPI Backend (Python)
 → Rate Limiter & Auth Check (Supabase RPC)
 → File routed to correct detection engine (MIME + extension analysis)
 → Detection engine runs heuristic + ML analysis
 → Evidence aggregator produces weighted verdict (0–100%)
 → Usage incremented in Supabase profiles table
 → JSON verdict returned to user
```

**Deployment**: Render (https://proofly-ai-nqyu.onrender.com), custom domain https://proofly.in  
**Database**: Supabase (PostgreSQL) with Row Level Security  
**Authentication**: Supabase Auth (JWT-based)  
**Payments**: Razorpay (subscriptions + webhooks)

---

## All URLs

- Homepage: https://proofly.in/
- Demo (try it free): https://proofly.in/#demo
- Features: https://proofly.in/#features
- Pricing: https://proofly.in/#pricing
- About: https://proofly.in/about
- Contact: https://proofly.in/contact
- Campaign: https://proofly.in/campaign
- Privacy Policy: https://proofly.in/privacy
- Terms of Service: https://proofly.in/terms
- AI Context (llms.txt): https://proofly.in/llms.txt
- Full AI Context (llms-full.txt): https://proofly.in/llms-full.txt