10 AI AGENTS · LIVE INFERENCE · NOTHING MOCKED

An AI that
fights AI.

Misinformation in India spreads through WhatsApp forwards, coordinated bot networks, and AI-generated images — in Hindi, Hinglish, and English. ShadowTrace hunts all three, in real time.

GROQ LLAMA-3.3-70BNEO4J GRAPHELA + 3-MODEL ENSEMBLESARVAM AI
LIVE SIMULATIONAUTHENTIC ACCOUNTAMPLIFYING FALSEHOODDETECTED BY SHADOWTRACE
01 · THE THREAT

Falsehood outruns the fact-check. Every time.

By the time a claim is manually debunked, it has already reached millions of family group chats. Detection has to happen at the speed of the campaign — not the newsroom.

0%
of Indians encountered fake news online
MICROSOFT DIGITAL CIVILITY
0M+
WhatsApp users — misinformation’s fastest vector
META, INDIA
0x
faster than truth — how falsehoods spread
MIT, SCIENCE
02 · THE SYSTEM

Ten specialised agents. One investigation.

Every agent runs real inference on input you provide — no canned results, no staged datasets. Type an account, paste a claim, drop an image URL.

AGENTS 01 · 06

Content & WhatsApp Intel

Groq Llama-3.3-70B scores any claim for misinformation, blended with forward-chain pattern detection built for Hindi, Hinglish, and English.

LIVE LLM INFERENCE
AGENT 02

Deepfake & Image Forensics

Error Level Analysis heatmaps plus a three-model classifier ensemble. An image is only called AI-generated when two models independently agree.

3-MODEL ENSEMBLE
AGENTS 07 · 09

Coordination Detection

Enter any real account. Posts are ingested live, then analysed for 60-second posting sync, stylometric fingerprints, and LLM-generation signals.

LIVE ACCOUNT INGESTION
AGENT 03

Network Graph Mapping

Neo4j-backed campaign topology. Origin nodes, bot clusters, and amplifier chains rendered as an explorable force-directed graph.

NEO4J AURADB
AGENT 10

Indian Language Detection

Sarvam AI identifies code-mixed and regional text — because misinformation in India does not arrive in English.

10+ INDIAN LANGUAGES
AGENTS 04 · 05

Threat Classification

A LangGraph pipeline chains every signal into one verdict: organic misinformation, coordinated inauthentic behaviour, or state-level operation.

LANGGRAPH ORCHESTRATION
03 · THE INVESTIGATION

One click. Five agents. A verdict you can audit.

Paste a WhatsApp forward and watch the chain execute — each agent reporting its own finding, with real latency, ending in an LLM-written threat assessment.

INPUT ▸ “Doctors at AIIMS confirmed hot lemon water kills corona virus. Share before deleted!!”
01WhatsAppAnalyzerForward chain detected · 4 red flags
02ContentAnalyzerGroq LLM misinformation score: 86/100
03SarvamLanguageDetectorLanguage: Hinglish · 100% confidence
04FactCheckCrossRefCross-referenced live fact-checker feeds
05ThreatClassifierhealth_misinformation · severity HIGH
THREAT ALERTHealth MisinformationSEVERITY: HIGH · SCORE 89/100

The claim asserts a home remedy cures a viral infection — false, and dangerous if it displaces medical treatment. Written by the model at inference time.

FASTAPILANGGRAPHGROQ LLAMA-3.3-70BNEO4J AURADBSARVAM AIHUGGING FACESCIKIT-LEARNNEXT.JSD3.JSBLUESKY APIPILLOW ELAFASTAPILANGGRAPHGROQ LLAMA-3.3-70BNEO4J AURADBSARVAM AIHUGGING FACESCIKIT-LEARNNEXT.JSD3.JSBLUESKY APIPILLOW ELA

Run it yourself.

Paste any claim. Type any real account. Drop any image URL. The dashboard is live — every score you see is computed the moment you ask for it.

ACCESS MISSION CONTROL →