PATAS
Pattern-Adaptive Transmodal Anti-Spam System
An offline pattern discovery engine that turns historical chaos into transparent, production-grade signals.
What is PATAS?
PATAS is a signal engine, not an enforcement system. It sits alongside your pipeline, analyzing history to inform the future.
Historical Logs
Analyzes message batches (10k-50k+) offline.
Signal Engine
Pattern Discovery & Rule Management
Transparent Rules
Outputs SQL-like rules & metrics for your existing engine.
The Problem
The Scope
"Given historical logs, automatically discover patterns, generate transparent rules, and evaluate them offline with clear safety thresholds."
Two-Stage Pattern Mining
Optimized for cost and precision
Stage 1: Fast Scan
Deterministic & Cheap
URLs, Domains, Keywords
~2-4 mins / 500k msgs
Stage 2: Deep Analysis
Semantic & Focused
Embeddings + Clustering + LLM
70-90% Cost Reduction
Rule Lifecycle Management
From discovery to enforcement with safety guarantees
Candidate
Generated, not evaluated
Shadow
Evaluated on history, inactive
Active
Safe for production enforcement
Deprecated
Disabled due to low precision
Conservative Profile
Recommended for high-risk automated blocking
Balanced Profile
For shadow mode & controlled experiments
Architecture Snapshot
Layered design for on-premise deployment
API Layer
FastAPI (app/api/)
Service Layer
Business Logic (app/v2_*.py)
Repository Layer
Data Access (SQLAlchemy 2.0)
Internal Benchmarks
Pattern Mining Speed (500k msgs)
*Benchmarks on 500k message dataset. Intended for offline overnight runs.
Pilot Execution Plan
Data Selection
Select 7-30 days of logs for a specific region.
Deployment
Deploy PATAS on-prem with read-only access.
Mining Run
Execute the two-stage pipeline offline.
Review & Export
Review metrics, export 'SAFE_AUTO' rules.
Maintenance
Weekly runs for new patterns & deprecation.