System Architecture

Autonomous Persistent Multi-Agent Cyber Threat Intelligence / V1.5
Minas / Core / Trust Zone 1
Maximum Trust / Orchestration, API, Processing
mandos
Orchestrator
LangGraph · FastAPI · Intel processing
tirith
Analyst Console
React app · NL query interface
annun
Source Discovery
4-tier engine · HITL approval
fea
Redis HA
Investigation state & checkpoints
ladris
Neo4j
Relationship graph of actors & campaigns
PostgreSQL
Primary Database
Evidence, threats, customer profiles · PGVector
Object Storage
Evidence Vault
Immutable · content-addressed
Trusted core where investigations are orchestrated and evidence is stored
Ithil / Enrichment / Trust Zone 2
Mediated Trust / External Tool Execution
telchar
Tool Gateway
MCP · external execution
ithildin
OSINT Collectors
Shodan · VirusTotal · WhoisXML
Controlled zone for external tool execution and OSINT collection
Mordor / Hostile / Trust Zone 3
Untrusted / Dark Web Collection
shelob
Dark Web Collector
Ephemeral Tor containers · no persistent state
Containers created per-crawl, destroyed afterward
Istar / Inference / GPU Isolated
Core Trust / Dedicated GPU Inference
halbarad
Qwen 2.5-14B
Triage & extraction
lorin
Llama 3.1 70B
Synthesis & escalation
Anthropic
Claude Opus 4.7
Reasoning & investigations
Hybrid 60/40 local-to-frontier routing with continuous quality monitoring and auto-failover to Claude on drift
Amon / Observability
Core Trust / Monitoring & Audit
sul
Observability Stack
Langfuse · Grafana · Prometheus
Full-system tracing, metrics, and quality monitoring across the platform

Five-Tier Storage Model

Tier 1 / Redis HA
Active investigation state
LangGraph checkpoints
Social deduplication
Tier 2 / PostgreSQL
Evidence / Threat intelligence
Customer profiles / Config
PGVector semantic search
Tier 3 / Neo4j
Relationship graph
Actors, campaigns, TTPs
Infrastructure linkage
Tier 4 / Object Storage
Immutable evidence vault
Content-addressed
Long-term retention
Tier 5 / Langfuse
Audit & observability
LLM traces / Cost tracking
Routing quality metrics

Architecture Principles

01Investigations are long-lived objects
02Evidence is immutable & content-addressed
03Tiered autonomy with HITL gates
04Hostile content assumed on all inputs
05Three trust zones enforced at network level
0660/40 hybrid LLM routing (local-first)
07Full HA from day one
08Customer digital twin model
09Quality governance & evidence grounding

Source Categories

Category 01 / Authorized OSINT Shodan / VirusTotal / WhoisXML
CISA KEV / NVD
Enrichment Zone Collectors
Category 02 / Social & Open Web Telegram / Mastodon / Bluesky
Publication RSS/Atom feeds
Researcher credibility tiers
Core Zone Collectors
Category 03 / Dark Web Approved .onion sites
Ransomware leak sites & forums
Ephemeral Tor containers
Hostile Zone Collectors