Poetry Intelligence System
A continuously learning agency intelligence platform with three differentiated layers: Causal AI, Temporal Knowledge Graph, and Cross-Client Pattern Learning.
Three-Layer Intelligence Architecture
AI-forward agency intelligence with causal validation and cross-client learning
🔬 Causal AI Layer
The "WHY" - Proving causation, not just correlation
- ✓ DoWhy/EconML integration
- ✓ Lift validation on MMM
- ✓ "Validated ROI" badge
- ✓ Geo Holdout testing
🧠 Temporal Memory
The "WHAT" - Zep/Graphiti bitemporal pattern
- ✓ Event + ingestion time
- ✓ Semantic entity subgraph
- ✓ Community clustering
- ✓ Point-in-time queries
🌐 Cross-Client Learning
The "HOW" - Network effect moat
- ✓ Privacy-safe aggregation
- ✓ Vertical pattern libraries
- ✓ k-anonymity (min 5 clients)
- ✓ No individual data exposed
Competitive Advantage
Poetry's moat is NOT any single AI capability but the integration of three layers that competitors struggle to replicate: Causal AI (the "why"), Temporal Knowledge Graph (the "what"), and Cross-Client Learning (the "how").
Comprehensive Signal Taxonomy
8 signal categories with real-time to quarterly capture frequencies
Tier 1: Built (USE NOW)
- ✓ Meta Ads (9 workers)
- ✓ Google Ads (8 workers)
- ✓ LinkedIn ingestion
Tier 2: n8n (QUICK WIN)
- ✓ GA4 signal ingest
- ✓ News/RSS monitor
- ✓ Slack alerting
Tier 3: Custom Workers
- ▢ Meta Ad Library
- ▢ Pathmatics/Vivvix
- ▢ Brandwatch
Causal AI & Real-time Anomaly Detection
DoWhy/EconML causal inference + 67% faster anomaly detection
🔬 Causal Validator Function
- Geo Holdout (highest confidence)
- Synthetic Control (medium)
- RCT when available
🚨 Anomaly Detector Function
Signal Aggregation Workflow
Daily 6AM trigger → Parallel n8n collection (GA4, News, Media, Competitive) → Anomaly detection → Slack alerts
signal-aggregation.bpmn
Anomaly Response Workflow
Severity routing (Critical: 1h, High: 4h SLA) → Causal analysis → Human-in-loop investigation → Resolution
anomaly-response.bpmn
McKinsey Finding: 40% of ad spend is wasted without causal validation. Poetry's integration of DoWhy/EconML with the MMM Allocator provides "Validated ROI" badges on recommendations.
Cross-Client Learning Patterns
Privacy-safe aggregate patterns creating a network effect moat
🎬 Pharma CTV Fatigue Pattern
🎮 Gaming Channel Synergy Pattern
Pattern Types
Privacy Architecture
Privacy-safe aggregate learning with k-anonymity guarantees
🔒 Privacy Guarantees
- ✓ k-anonymity (minimum 5 clients per pattern)
- ✓ No client-level data ever exposed
- ✓ Aggregate-only pattern storage
- ✓ Differential privacy layer
- ✓ Client cannot be re-identified from patterns
- ✓ Data never leaves client scope
How Cross-Client Learning Works
Privacy vs Value Trade-off
Never shared, never exposed, stays in client scope
Only aggregate metrics, stripped of identifiers
Insights from 100+ patterns benefit all clients
Implementation Components
Knowledge Graph schemas, CortexOne functions, and BPMN workflows
Neo4j Schemas (4)
signals-intelligence.cyphertemporal-memory.cyphercausal-validation.cyphercross-client-patterns.cypher
Location: packages/knowledge-base/poetry/schema/
CortexOne Functions (4)
- Causal Validator
- Anomaly Detector
- Pattern Matcher
- Competitive Monitor
Location: cortexone-functions/
BPMN Workflows (2)
signal-aggregation.bpmnanomaly-response.bpmn
Location: packages/bpmn/poetry/processes/
⚡ n8n Signal Ingestion Workflows (3)
Google Analytics 4 → Knowledge Base
News/RSS → Competitive intelligence
Anomaly alerts → Slack notifications
Location: packages/n8n/workflows/
Key Takeaways
What makes Poetry Intelligence System a competitive moat
Causal Validation
DoWhy/EconML integration proves causation, not just correlation. 40% of ad spend identified as wasted without causal validation.
Temporal Memory
Zep/Graphiti-style bitemporal knowledge graph enables point-in-time queries and 18.5% accuracy improvement.
Network Effect Moat
Cross-client learning with privacy guarantees creates value that increases with each new client.
Privacy-First Design
k-anonymity (min 5 clients), differential privacy, and aggregate-only patterns ensure client data protection.
Three-Layer Integration
The "WHY"
Validates causation
The "WHAT"
Bitemporal context
The "HOW"
Network effect moat
Schemas: packages/knowledge-base/poetry/schema/ | Functions: cortexone-functions/ | n8n: packages/n8n/workflows/