LangGraph + CrewAI + AutoGen + Ollama

TruthLens DeepFake Detection

3-Triple Multi-Agentic AI Framework
React Dashboard · Real-time BI · ADAG Red Team

89/89 Tests PASS
7 Detection Modules
14 Dashboard Pages
45 REST APIs

Core Differentiators

TruthLens v2.8.0 - 6 Core Strengths

💜

Biological Signal Verification

rPPG heart rate extraction and eye blink pattern analysis detect physiologically impossible patterns - the strongest evidence against deepfakes.

Most reliable evidence
🤖

Triple Agentic AI

LangGraph orchestration + CrewAI 6-expert consensus + AutoGen adversarial debate - three frameworks working in concert.

3 AI Frameworks
🔒

Privacy Guaranteed

Ollama-based local LLM/VLM inference ensures zero data transmission to external servers. Safe for sensitive data.

100% Local Processing
💡

Explainable AI (XAI)

Every verdict includes LLM-generated detailed explanation and evidence chain for full transparency.

Evidence Chain
🛡️

ADAG Red Team

Adversarial testing framework with 4 attack modules automatically verifies detection robustness via Red-Blue Team feedback loop.

4 Attack Modules

Real-time Processing

Send() parallel dispatch, async architecture, MemorySaver checkpoint for fault-tolerant real-time analysis.

60% faster

Triple Agentic AI

LangGraph + CrewAI + AutoGen - 3 frameworks in synergy

1

LangGraph — Workflow Orchestration

17-node async StateGraph pipeline with Send() parallel dispatch for 5 analysis nodes running simultaneously. MemorySaver checkpoint enables fault recovery. Human-in-the-Loop (HITL) for uncertain zone review.

Nodes 17
Parallel 5 agents
Speedup 60%
Preprocess
Visual
Audio
Bio
OCR
A/V Sync
Fusion
XAI + HITL
Verdict
2

CrewAI — 6-Expert Consensus Panel

👁️ Visual Forensics
🎵 Audio Forensics
💜 Biometric
📝 OCR/Document
👄 A/V Sync
👑 Chief Analyst

6 specialized expert agents analyze evidence independently, then the Chief Analyst synthesizes all findings for the final verdict. Supports Sequential and Hierarchical process modes with SQLite memory backend for learning from past judgments.

Experts 6
Memory SQLite
Model llama3.2:3b
3

AutoGen — Adversarial Debate System

When fusion scores fall in the uncertain zone (0.35~0.65), adversarial debate automatically triggers. Prosecutor argues FAKE, Defense argues REAL, and Judge renders the final verdict. Supports 1v1 and 6-person GroupChat modes.

Trigger 0.35~0.65
Avg Time ~19s
Modes 1v1 / Group
⚔️ Prosecutor FAKE
VS
🛡️ Defense REAL
⚖️ Judge VERDICT

React Dashboard

14 pages · Real-time BI · WebSocket · 45 REST APIs

Media Analysis

☁️

Drag & Drop media files here

Video (MP4, AVI, MOV) · Image (JPG, PNG) · Audio (MP3, WAV)

Max 500MB

📹 Video
🖼️ Image
🎵 Audio
🌐 URL

BI Dashboard

Auto-refresh: 10s
1,247
Total Analyses
342
Fakes Detected
6.3s
Avg Response
92.5%
Accuracy
Daily Requests (30 days)
Verdict Distribution
REAL 58%
FAKE 27%
UNCERTAIN 15%

Analysis History

DateTypeVerdictConfidenceTime
2026-03-06 14:23videoFAKE85.3%6.2s
2026-03-06 13:45imageREAL92.1%3.8s
2026-03-06 12:10videoUNCERTAIN54.7%8.1s
2026-03-05 18:30audioREAL88.9%4.5s

Agent Settings

LangGraph Pipeline Orchestration

Enable Checkpoint
Human Review (HITL)
Review Threshold
0.40

Active Nodes (12)

Preprocess Visual Audio Biological Fusion AutoGen XAI Verdict C2PA SynthID Forensic Log Report

AutoGen Debate System

Debate Configuration

Modelllama3.2:3b
Max Rounds3
Mode1v1
Threshold0.35 ~ 0.65
Always Debate

LLM / Ollama Settings

Server URLhttp://localhost:11434
StatusConnected

Model Assignments

LLM Modelllama3.2:3b
VLM Modelllama3.2-vision
Embeddingnomic-embed-text

Performance

Temperature0.10
Context Window4096

ADAG Red Team

Biological Signal Injector

Fake rPPG + blink pattern injection

GAN Fingerprint Disruptor

FFT masking + camera ISP simulation

Temporal Manipulator

Micro-motion + motion blur injection

Text Humanizer

Statistical noise + typo injection

47 Total Tests
23.4% Evasion Rate
Visual Most Vulnerable

User Management

NameEmailRole
Brian Leebrian@a3sec.comADMIN
Analyst Kimkim@a3sec.comANALYST
Viewer Parkpark@a3sec.comVIEWER

Organizations

A3 SecurityENTERPRISE
Monthly Quota10,000
Usage

API Keys

API Keytl_sk_****...****3f7a
Created2026-02-15

Settings

Language

KO EN JA ZH

Profile

NameBrian Lee
RoleADMIN

API Documentation

SDK Examples

Python TypeScript cURL
import requests

resp = requests.post(
    "http://localhost:7000/ai/detect/upload",
    files={"file": open("video.mp4","rb")},
    headers={"X-API-Key": "tl_sk_..."}
)
print(resp.json()["verdict"])

Endpoints (45)

POST /ai/detect/upload
POST /ai/detect/url
GET /ai/detect/status/{id}
GET /ai/detect/analytics
GET /ai/usage/stats

7 Detection Modules

Multimodal analysis with weighted fusion

👁️

Visual Forensics

GAN fingerprint, compression artifacts, face manipulation

Weight: 25%
VLMFFTOpenCV
🎵

Audio Analysis

TTS/voice cloning detection, spectrum analysis

Weight: 20%
MFCCZCRLibrosa
💜

Biological Signal

rPPG heart rate extraction, eye blink pattern analysis

Weight: 30% (Highest)
MediaPipeCHROMHeartPy
👄

A/V Sync

Lip-sync verification, speaker consistency

Weight: 15%
ResemblyzerMediaPipe
📝

OCR/Document

Subtitle manipulation, document forgery, text consistency

Weight: 10%
PaddleOCR
🔄

Cross-Modal

Cross-modal anomaly detection via Whisper transcription

Auxiliary
Whisper
🧠

Few-Shot

Prototypical Networks for novel deepfake type learning

Auxiliary
Proto-Net

Dynamic Weighted Fusion

final_score = (Visual × 0.25 + Audio × 0.20 + Bio × 0.30 + AV_Sync × 0.15 + OCR × 0.10) / Σweights
REAL ≤ 0.35
0.35 < UNCERTAIN < 0.65
FAKE ≥ 0.65

Architecture & Tech Stack

Production-ready system with 89/89 tests passing

Frontend
React / Vite TanStack Query Recharts WebSocket
Backend
FastAPI 45 REST APIs 2 WebSocket yt-dlp
AI Core
LangGraph CrewAI AutoGen Ollama
ML/CV
MediaPipe OpenCV Librosa HeartPy PaddleOCR Whisper

Core Technologies

🔄 LangGraph 17-Node StateGraph
🤖 CrewAI 6-Agent Consensus
💬 AutoGen Adversarial Debate
🦙 Ollama Local LLM/VLM
⚛️ React / Vite Dashboard UI
FastAPI REST + WebSocket
0
/89
Tests PASS
0
modules
Detection Modules
0
pages
Dashboard Pages
0
APIs
REST Endpoints
0
frameworks
Agent Frameworks
0
%
Local Processing