Comparison and Positioning

PATAS: Comparison & Positioning

This document provides a detailed comparison of PATAS with traditional approaches and existing solutions in the anti-spam/abuse detection market.

---

PATAS vs Traditional Approaches

PATAS vs Classic ML Classifiers

Aspect	Classic ML Classifier	PATAS
Pattern Discovery	Requires labeled training data, manual feature engineering	Automatic discovery from historical logs
Transparency	Black-box model (neural network, ensemble)	Transparent SQL/regex rules
Explainability	Hard to explain why a message was flagged	Clear: shows exact rule and why it matched
Metrics	Global model metrics (accuracy, F1)	Per-rule metrics (precision, recall, coverage)
Rule Management	Retrain entire model	Individual rule lifecycle (promote/deprecate)
Safety Control	Single threshold for all decisions	Tiered safety profiles (Conservative/Balanced/Aggressive)
On-Premise	Often requires cloud ML services	Fully on-premise, no external dependencies

PATAS vs Classic LLM Filters

Aspect	Classic LLM Filter	PATAS
LLM Usage	Real-time per-message classification (online)	Pattern discovery from aggregated data (offline)
Pattern Discovery	LLM analyzes individual messages in real-time	Automatic clustering + LLM analyzes aggregated patterns
Rule Generation	LLM suggests rules per message (may be inconsistent)	LLM discovers patterns, then deterministic SQL/regex generation
Cost	Per-message API calls (expensive at scale)	One-time pattern discovery on aggregated data, then rule-based
Latency	API call per message (100-500ms)	Rule evaluation (1-10ms)
Transparency	LLM reasoning may be opaque	Transparent SQL rules (LLM used for discovery, not classification)
Consistency	LLM may give different answers for similar messages	Deterministic rule matching (LLM discovers patterns once)
Privacy	Messages sent to external API in real-time	Aggregated data only, encrypted/PII-redacted, on-premise option
Data Volume	Processes each message individually	Processes aggregated patterns, scales to millions

Key Difference: Classic LLM Filters use LLM for real-time classification of individual messages. PATAS uses LLM for offline pattern discovery from aggregated data, then uses deterministic rules for real-time evaluation.

PATAS vs Manual Rule Writing

Aspect	Manual Rule Writing	PATAS
Discovery	Human analysts identify patterns	Automatic discovery from data (with LLM semantic understanding)
Coverage	Limited by analyst time/expertise	Discovers patterns at scale
Maintenance	Manual updates when spam evolves	Automatic re-discovery
Metrics	Often no per-rule metrics	Built-in precision/recall/coverage
Safety	Ad-hoc safety decisions	Systematic safety profiles
Scalability	Doesn't scale with data volume	Scales automatically

---

PATAS vs Market Solutions

Comparison Table

Feature	Sardine	Abnormal	Feedzai	Stream	PATAS
Auto pattern discovery	❌	✅ (LLM)	❌	❌	✅ (data-driven + LLM)
Transparent rules	✅	~	✅	✅	✅
Per-rule metrics	✅	❌	~	❌	✅
Safety profiles	❌	❌	❌	❌	✅
On-premise	~	❌	~	❌	✅
General-purpose	❌ (fintech)	❌ (email)	❌ (fintech)	✅	✅
Developer API-first	✅	~	~	❌	✅

Detailed Comparison

Sardine

• Focus: Fintech fraud detection
• Strengths: Transparent rules, per-rule metrics
• Limitations: No automatic pattern discovery, fintech-specific, limited on-premise
• PATAS Advantage: Automatic discovery, general-purpose, full on-premise

Abnormal Security

• Focus: Email security
• Strengths: LLM-based pattern discovery
• Limitations: Email-specific, SaaS-only, no per-rule metrics, no safety profiles
• PATAS Advantage: Data-driven discovery (not LLM-dependent), general-purpose, on-premise, per-rule metrics

Feedzai

• Focus: Fintech fraud prevention
• Strengths: Transparent rules, some on-premise options
• Limitations: Fintech-specific, no automatic discovery, limited per-rule metrics
• PATAS Advantage: Automatic discovery, general-purpose, comprehensive per-rule metrics

Stream

• Focus: General-purpose content moderation
• Strengths: General-purpose, transparent rules
• Limitations: No automatic discovery, SaaS-only, no per-rule metrics, no safety profiles
• PATAS Advantage: Automatic discovery, on-premise, per-rule metrics, safety profiles

---

What Makes PATAS Unique

No other product combines all of these features:

1. ✅ Automatic pattern discovery from historical logs (not manual rule writing)
2. ✅ Transparent rule generation (SQL/regex-like, not black-box models)
3. ✅ Per-rule metrics (precision, recall, coverage for each rule)
4. ✅ Safety profiles (Conservative/Balanced/Aggressive tiers)
5. ✅ On-premise / self-hosted option (not SaaS-only)
6. ✅ General-purpose (not limited to email or fintech)
7. ✅ Developer API-first (not just UI for non-technical teams)

---

When to Use PATAS

PATAS is ideal when you need:

• Transparency: You need to explain why messages were flagged
• Control: You want per-rule metrics and safety profiles
• Privacy: You need on-premise processing
• Scale: You have large volumes of historical logs to analyze
• Flexibility: You need a general-purpose solution (not domain-specific)
• Developer-friendly: You prefer API-first over UI-only tools

PATAS may not be ideal if:

• You need real-time per-message LLM analysis (use LLM filters for real-time classification)
• You prefer black-box ML models (use traditional ML classifiers)
• You only need simple keyword blocking (use basic regex rules)
• You require a SaaS-only solution (PATAS is on-premise-first)

---

Integration with Existing Systems

PATAS is designed to complement, not replace, existing anti-spam systems:

• Use PATAS for pattern discovery and rule generation (offline analysis)
• Use your existing ML/LLM for real-time classification if needed
• Combine signals from both systems for final decisions
• Use PATAS metrics to validate and improve your ML models

This hybrid approach gives you the best of both worlds: automatic discovery with semantic understanding (PATAS) + real-time classification (ML/LLM filters).

---

Last Updated: 2025-11-18