Warcast

WARCAST

WAR ROOM

How Warcast Works

Real-time global conflict intelligence powered by custom ML models and AI orchestration

Our Mission

Warcast monitors global conflicts in real-time by aggregating intelligence from 40+ international news sources, ACLED verified conflict data, GDELT event streams, and social media. Every article is scored by our custom-trained ML model running on-device, with Anthropic Claude providing context-aware scoring via RAG. Our War Agent — also powered by Claude — delivers interactive threat analysis, 48-hour predictions, and multi-agent simulations to help you stay ahead of escalating tensions worldwide.

Custom ML Scoring Engine

WarCast Score (WC)

Every article receives a WarCast Score from our locally-trained machine learning model — no API call required. The local ML model handles fast preprocessing and serves as a fallback score.

  • • Model: GradientBoostingRegressor with 431-dimensional feature vectors
  • • Embeddings: Sentence-transformers (all-MiniLM-L6-v2) for semantic text understanding
  • • Structured Features: Military keyword detection, major power involvement, weapon mentions, geographic hotspots
  • • Incremental Learning: Model improves from admin corrections (weighted 10x) and domain heuristics
  • • Uncertainty Tracking: Each prediction includes a confidence estimate based on recent error patterns

Brain Score (Anthropic Claude)

The Brain engine uses RAG (Retrieval-Augmented Generation) to retrieve relevant active conflicts, recent analyses, and conflict memory, then scores each article with full situational awareness via Anthropic Claude. This is the authoritative score displayed across the platform. Claude Haiku handles bulk scoring; Claude Sonnet handles deep analysis, hourly briefs, and alert assessment.

Global Threat Level (GTL)

The GTL is computed locally every 5 minutes using a weighted formula across all recent articles, ACLED events, and social media signals. It powers the doomsday clock on the homepage and drives the threat distribution across Critical (≥90), High (≥50), Moderate (≥30), and Low (<30) columns.

War Agent Intelligence System

Interactive AI Analysis

The War Agent is powered by Anthropic Claude with native tool-use capabilities, enabling it to actively search our news database, query the web, retrieve actor profiles, and pull verified ACLED conflict timelines during conversation:

  • • News Search: Queries Warcast's database of 17,000+ articles for relevant intelligence
  • • Web Search: Self-hosted search engine for real-time supplementary context
  • • Actor Profiles: Detailed capability and decision-pattern profiles for major state actors
  • • Conflict Timelines: Verified ACLED event data for historical pattern matching

48-Hour Predictive Forecasts

Select any high-impact event to generate a multi-scenario forecast:

  • • Escalation Probability: Calculated 24h, 48h, and 1-week risk assessments
  • • Key Decision Points: Identifies critical moments and alternative outcomes
  • • Actor Behavior Modeling: Predicts responses from USA, Russia, China, NATO, Iran based on historical patterns
  • • Early Warning Signals: Five signal types — escalation, pattern match, anomaly, threshold, correlation

Multi-Agent Simulation

Each major actor has custom decision profiles built from historical behavior patterns. The simulation models diplomatic, economic, and military responses across multiple actors simultaneously, generating outcome probabilities for each scenario.

Persistent Chat History

All War Agent conversations are stored with full context — resume any session, browse history, and share analysis. The agent is injected with live threat data, the latest hourly intelligence briefing, and current market conditions before each response.

Data Pipeline

1. Intelligence Collection

Continuous multi-source monitoring updated every 2 minutes:

  • • ACLED: Armed Conflict Location & Event Data — verified, ground-truth conflict events
  • • GDELT: Global Database of Events, Language, and Tone (2.0 API)
  • • X (Twitter): Real-time social media monitoring from conflict-focused accounts
  • • Web Search: Self-hosted search engine for breaking news and social post enrichment
  • • RSS Feeds: 40+ international news sources polled every 2 minutes
  • • Priority System: Sources ranked 1-10 by reliability, with automatic disable after repeated failures

2. Processing Pipeline

Every article passes through multiple scoring and analysis stages:

  • • Local ML Scoring: Custom GradientBoosting model scores articles in under 50ms
  • • Brain Scoring: Anthropic Claude scores articles with RAG context (active conflicts, prior analyses)
  • • Entity Extraction: Countries, organizations, weapons, military units, key figures
  • • Duplicate Detection: Multi-layer similarity scoring (title 50%, description 20%, keywords 30%)
  • • Social Post Enrichment: High-value tweets enriched with web search for related coverage

3. Threat Classification

  • • Critical (90-100): Nuclear threats, WMD deployment, declarations of war
  • • High (50-89): Invasions, major offensives, military operations, regional conflicts
  • • Moderate (30-49): Military exercises, border tensions, arms deals
  • • Multipliers: Major power involvement, casualty reports, nuclear keywords boost scores

4. Real-time Delivery

WebSocket-first architecture — no polling:

  • • Persistent Connections: All data flows through WebSocket v2 channels
  • • Instant Updates: New articles pushed to connected clients as they're scored
  • • Filter Subscriptions: Subscribe to specific sources, threat levels, or regions
  • • Redis Caching: Fast initial loads with invalidation on new data

Technical Architecture

Frontend

  • • Next.js 16 (App Router, Turbopack)
  • • React 19 with TypeScript
  • • Tailwind CSS
  • • React Query for state management
  • • Privy for authentication

Backend

  • • FastAPI (Python 3.12)
  • • SQLAlchemy 2.0 + Alembic migrations
  • • MariaDB + Redis
  • • AsyncIO task scheduler
  • • WebSocket manager v2

AI / ML

  • • Custom GradientBoosting model (scikit-learn)
  • • Sentence-transformers embeddings
  • • Anthropic Claude (Brain scoring + War Agent)
  • • Self-hosted web search engine

Infrastructure

  • • Systemd services on dedicated VPS
  • • Nginx reverse proxy
  • • Cloudflare CDN
  • • Solana/Jupiter integration
  • • Android app (Kotlin + Jetpack Compose)

Smart Alert System

Configurable alerts with per-user settings, source filtering, and cooldowns:

Alert Types

  • • Escalation warnings (confidence-based)
  • • Verified threat notifications
  • • Pattern match alerts
  • • Anomaly detection

Delivery

  • • In-app real-time notifications
  • • Telegram critical alerts
  • • Automated X (Twitter) briefs
  • • Configurable thresholds and regions

War Room Capabilities

Active Features

  • • War Agent Chat — interactive AI analysis powered by Anthropic Claude
  • • 48-Hour Forecasts — multi-scenario predictive analysis with escalation probabilities
  • • Multi-Agent Simulations — actor behavior modeling with custom decision profiles
  • • Smart Alerts V2 — per-user configs with source filtering and cooldowns
  • • Threat Tracker — real-time monitoring with ACLED/GDELT integration
  • • Market Intelligence — GTL vs BTC/S&P 500/Gold correlation analysis
  • • Entity Tracking — relationship mapping for countries, organizations, and key figures
  • • Social Media Intelligence — X (Twitter) monitoring with web search enrichment
  • • Automated X Briefs — hourly intelligence summaries posted to @WarCastApp

In Development

  • • Android app — native Kotlin app with real-time updates (Play Store)
  • • Push notifications — mobile and web alert delivery
  • • Multi-language support for global coverage

Technical Notes

Scoring Philosophy

The Brain Score (Anthropic Claude + RAG) is the authoritative metric. It scores each article with full situational awareness using retrieval-augmented generation over active conflicts, recent analyses, and conflict memory. Claude Haiku handles bulk scoring for efficiency; Claude Sonnet handles deep analysis, hourly intelligence briefs, and alert assessment. The local ML model handles fast preprocessing and learns incrementally from both Brain scores and admin corrections, serving as a fallback when the Brain is unavailable.

Performance

War Agent responses take 10-30 seconds due to real-time tool execution — the agent may search our database, query the web, and retrieve ACLED data before responding. We prioritize accuracy over speed for intelligence assessments. Local ML scoring is near-instant; Brain scoring via Anthropic Claude processes asynchronously in the background.

Data Privacy

Chat histories are stored securely and linked to user accounts. Anonymous users get a temporary ID. We don't sell or share data with third parties. API calls to Anthropic include only the necessary context for analysis, never personal information.

Back to Monitor