AI Engine Architecture
The Python AI Engine is the intelligence core of QBITEL Bridge, handling protocol discovery, threat analysis, and autonomous security operations.
Overview
The AI Engine is a modular Python application built with FastAPI and PyTorch. It provides REST and gRPC APIs, an autonomous multi-agent system, and a complete ML pipeline for protocol discovery and security analysis.
Module Structure
ai_engine/
core/ # Engine core, config, orchestrator
agents/ # Multi-agent orchestration system
discovery/ # Protocol discovery (PCFG, transformers)
detection/ # Field detection (BiLSTM-CRF)
anomaly/ # Anomaly detection (Isolation Forest, LSTM, VAE)
security/ # Zero-touch decision engine
compliance/ # Compliance automation (9 frameworks)
copilot/ # Protocol intelligence copilot
crypto/ # Post-quantum cryptography
llm/ # LLM service (Ollama, RAG, providers)
marketplace/ # Protocol marketplace (Stripe, S3)
legacy/ # Legacy System Whisperer (COBOL, JCL)
cloud_native/ # Service mesh, container security
api/ # REST (FastAPI) and gRPC APIs
models/ # ML model management and registry
monitoring/ # Metrics, logging, tracing
tests/ # Test suite (unit, integration, perf) Protocol Discovery Pipeline
The discovery pipeline transforms raw network traffic into structured protocol definitions through six stages:
- Statistical Analysis -- byte distributions, entropy calculations, message length profiling
- Pattern Extraction -- delimiter detection, header identification, field boundary analysis
- PCFG Inference -- learn a Probabilistic Context-Free Grammar from traffic patterns
- Parser Generation -- auto-generate parsers from the learned grammar
- ML Classification -- classify using ensemble models (CNN, LSTM, Random Forest)
- Validation -- verify discovered protocols with structural and semantic checks
Key Discovery Modules
| Module | Responsibility |
|---|---|
statistical_analyzer | Entropy, byte frequency, n-gram analysis |
pattern_extractor | Recurring patterns, delimiters, headers |
enhanced_pcfg_inference | Grammar learning with production probabilities |
enhanced_parser_generator | Dynamic parser code generation |
protocol_classifier | Ensemble ML classification |
Multi-Agent System
The AI Engine implements a multi-agent orchestration framework where specialized agents collaborate autonomously:
- Agent Pool -- manages agent lifecycle, resource allocation, and health monitoring
- Agent Communication -- message-passing infrastructure between agents
- Agent Memory -- persistent and working memory for context retention
- Agent Collaboration -- task decomposition and result aggregation
- Planning Agent -- breaks complex tasks into executable sub-goals
Zero-Touch Decision Engine
The Zero-Touch Decision Agent provides autonomous security response with configurable confidence thresholds:
| Confidence | Risk Level | Action |
|---|---|---|
| ≥ 0.95 | Low | Autonomous execution |
| ≥ 0.85 | Medium | Auto-approved with logging |
| ≥ 0.50 | Any | Escalated to operator |
| < 0.50 | Any | Escalated to senior analyst |
LLM Integration
The AI Engine integrates with on-premise LLM providers for air-gapped deployments:
- Ollama -- primary provider for local LLM inference
- vLLM -- high-throughput serving for production workloads
- RAG Engine -- retrieval-augmented generation with Qdrant vector store
- Semantic Cache -- caches LLM responses for repeated queries
- Guardrails -- input/output validation for LLM safety
Anomaly Detection
An ensemble of anomaly detectors provides multi-layered threat detection:
- Isolation Forest -- unsupervised anomaly scoring
- LSTM Detector -- temporal sequence anomaly detection
- VAE Detector -- variational autoencoder for reconstruction-based detection
- Ensemble Detector -- weighted combination of all detector outputs
Next Steps
- REST API Reference -- explore the AI Engine's API endpoints
- Python Development Guide -- contribute to the AI Engine
- Rust Data Plane -- understand the data pipeline