VERITAS
— AI Counter-Disinformation Platform
Real-time disinformation detection, attribution, and counter-narrative generation for democratic governments
TIER 1 — DATA SOURCES
TIER 2 — COLLECTION
TIER 3 — MESSAGE BUS
TIER 4 — AI ANALYSIS PIPELINE
TIER 5 — KNOWLEDGE STORE
TIER 6 — PRODUCTION
TIER 7 — DELIVERY
CLASSIFIED PROCESSING BOUNDARY
APACHE KAFKA EVENT MESH
Telegram Channels & Groups
X/Twitter Firehose
TikTok Trends
YouTube Transcripts
News Wire APIs
Reuters, AFP, DPA
Dark Web / Telegram Deep
Onion / private channels
Multi-Platform Ingestion
Engine (Telegram MTProto, X API)
Real-time Stream Processor
Flink / Kafka Streams
Media Transcriber
Whisper → text, image OCR
Language Detector & Normalizer
fastText / CLD3 / transliteration
Narrative Extraction LLM
fine-tuned 70B / Mixtral 8x22B
Bot Coordination Detector
temporal burst + graph clustering
State Actor Attribution Engine
TTP fingerprint matching
Linguistic Forensic Analyzer
syntax + idiolect signatures
Cross-Cultural Context Model
Turkish / Islamosphere nuances
religious code, honor/shame
Factuality Scorer
claims vs verified ground truth
Campaign Knowledge Graph
Neo4j — actors, narratives, channels
Evidence Vault
immutable, chain-of-custody
Fingerprint Database
known campaign signatures
Real-time Threat Dashboard
live campaign topology
Automated Intelligence Reports
STIX 2.1 / MISP format
Counter-Narrative Generator
audience-tailored rebuttals
Early Warning Alert System
threshold-based + anomaly
Government Secure Portal
mTLS + SSO (SAML/OIDC)
REST API for Platform Integration
rate-limited, scoped keys
Air-Gapped Classified Variant
offline LLM + T5-tier hardware
secure channel
fingerprints
LEGEND
External Data Sources
Collection / Ingestion
AI / ML Processing
Detection / Attribution
Knowledge Store
Delivery / API
Message Bus
auth flow
Detection Pipeline
Raw content is normalized, transcribed, and language-
tagged. Burst detection identifies coordinated posting;
LLMs extract narratives and score factuality against
ground-truth databases. Anomaly alerts fire within 90s.
Attribution Methodology
Linguistic forensics + TTP fingerprints + infrastructure
overlap build an evidence chain. The Cross-Cultural
Context Model flags native-speaker markers that Western
tools miss. Confidence scores are expressed per STIX 2.1.
Deployment Models
1. SaaS (EU-hosted, GDPR)
2. On-premise government stack
3. Air-gapped classified variant
All models export STIX/MISP and REST.
VERITAS ARCH v1.0