Skip to main content

Architecture Overview

This document provides a comprehensive overview of the AI Personality Drift Simulation system architecture.

System Overview​

The system is designed as a microservices architecture with the following key components:

┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐
│   Frontend      │    │   FastAPI       │    │   Simulation    │
│   Dashboard     │◄──►│   Application   │◄──►│   Engine        │
└─────────────────┘    └─────────────────┘    └─────────────────┘
                                │
                                ▼
┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐
│   Redis         │    │   Qdrant        │    │   Ollama        │
│   (Sessions)    │    │   (Memory)      │    │   (LLM)         │
└─────────────────┘    └─────────────────┘    └─────────────────┘

Core Components​

1. API Layer (src/api/)​

FastAPI Application - Main entry point for all external interactions

# Key components:
- app.py              # Application factory and configuration
- routes/             # API endpoint definitions
  ├── simulation.py   # Simulation control endpoints
  ├── data.py         # Data access endpoints
  ├── monitoring.py   # Health and status endpoints
  └── websocket.py    # Real-time communication

Key Features:

  • RESTful API with OpenAPI documentation
  • WebSocket support for real-time updates
  • CORS configuration for frontend integration
  • Background task management
  • Static file serving for frontend

2. Simulation Engine (src/services/simulation_engine.py)​

Core orchestration service that manages the entire simulation lifecycle

class SimulationEngine:
    - initialize_simulation()    # Setup personas and conditions
    - run_simulation()          # Main simulation loop
    - pause_simulation()        # Pause execution
    - resume_simulation()       # Resume execution
    - stop_simulation()         # Clean shutdown

Key Responsibilities:

  • Persona lifecycle management
  • Event scheduling and injection
  • Assessment administration
  • Time compression (5 years → 4-6 hours)
  • Progress tracking and checkpointing

3. Persona Management (src/services/persona_manager.py)​

Persona lifecycle and state management

class PersonaManager:
    - create_persona()          # Initialize new persona
    - update_persona_state()    # Update personality state
    - administer_assessment()   # Run psychiatric assessments
    - process_event()           # Handle simulation events
    - get_persona_memory()      # Retrieve memory data

Key Features:

  • Personality baseline establishment
  • State persistence and recovery
  • Memory integration with LLM responses
  • Cross-assessment consistency

4. Assessment System (src/assessment/)​

Psychiatric assessment administration and scoring

# Assessment scales:
- phq9.py           # Depression assessment (PHQ-9)
- gad7.py            # Anxiety assessment (GAD-7)
- pss10.py           # Stress assessment (PSS-10)
- assessment_service.py  # Orchestration service

Key Features:

  • Natural language question administration
  • Response parsing and validation
  • Clinical scoring and interpretation
  • Longitudinal trend analysis

5. Mechanistic Analysis (src/interpretability/)​

Neural circuit analysis during LLM inference

# Key components:
- attention_capture.py    # Attention weight extraction
- activation_patching.py  # Layer-wise interventions
- circuit_tracking.py     # Neural circuit monitoring
- drift_detection.py      # Change point detection

Key Features:

  • Real-time attention pattern capture
  • Activation patching for causal analysis
  • Self-reference circuit tracking
  • Drift detection algorithms

Data Flow​

1. Simulation Initialization​

sequenceDiagram
    participant API
    participant Engine
    participant PersonaManager
    participant LLM
    participant Storage

    API->>Engine: POST /simulation/start
    Engine->>PersonaManager: create_personas()
    PersonaManager->>LLM: establish_baseline()
    LLM->>Storage: store_baseline_data()
    Engine->>API: simulation_started

2. Event Processing​

sequenceDiagram
    participant Engine
    participant PersonaManager
    participant LLM
    participant Assessment
    participant Analysis

    Engine->>PersonaManager: inject_event(event)
    PersonaManager->>LLM: process_event()
    LLM->>Analysis: capture_mechanistic_data()
    PersonaManager->>Assessment: administer_assessment()
    Assessment->>Engine: store_results()

3. Assessment Administration​

sequenceDiagram
    participant Assessment
    participant LLM
    participant Parser
    participant Storage

    Assessment->>LLM: ask_question(question)
    LLM->>Assessment: response
    Assessment->>Parser: parse_response(response)
    Parser->>Assessment: score
    Assessment->>Storage: store_assessment(score)

Storage Architecture​

1. Redis (src/storage/redis_client.py)​

Session and cache storage

# Key data types:
- Simulation state          # Current simulation status
- Persona sessions         # Active persona data
- Assessment cache         # Cached assessment results
- Event queue             # Pending events

2. Qdrant (src/storage/qdrant_client.py)​

Vector database for memory storage

# Collections:
- persona_memories         # Long-term memory embeddings
- assessment_history       # Historical assessment data
- mechanistic_data        # Neural analysis results
- event_embeddings        # Event memory representations

3. File Storage (src/storage/file_storage.py)​

Data export and archival

# Export formats:
- CSV                     # Tabular data
- JSON                    # Structured data
- Parquet                 # Efficient binary format
- YAML                    # Configuration snapshots

Configuration Management​

1. Environment Configuration (src/core/config.py)​

Application-wide settings

class Settings:
    # Application settings
    APP_NAME: str
    VERSION: str
    API_PREFIX: str
    
    # Database connections
    REDIS_URL: str
    QDRANT_URL: str
    
    # LLM configuration
    OLLAMA_BASE_URL: str
    MODEL_NAME: str
    
    # Simulation parameters
    SIMULATION_TIMEOUT: int
    MAX_CONCURRENT_PERSONAS: int

2. Experiment Configuration (config/experiments/)​

Study-specific parameters

# clinical_thresholds.yaml
phq9:
  mild: 5
  moderate: 10
  severe: 15

# personality_drift.yaml
stress_scaling:
  low: 0.1
  medium: 0.3
  high: 0.5

# simulation_timing.yaml
compression_factor: 4380  # 5 years → 4 hours
assessment_interval: 7    # Weekly assessments

Security Considerations​

1. API Security​

  • CORS Configuration: Restricted origins
  • Input Validation: Pydantic model validation
  • Error Handling: Secure error responses

2. Data Security​

  • Environment Variables: Sensitive configuration

Performance Optimization​

1. LLM Optimization​

  • Model Quantization: Reduced memory footprint
  • Response Caching: Avoid redundant computations
  • Batch Processing: Efficient multi-persona handling
  • Memory Management: Garbage collection optimization

2. Simulation Optimization​

  • Time Compression: Accelerated simulation loops
  • Parallel Processing: Concurrent persona management
  • Checkpointing: Efficient state persistence
  • Resource Monitoring: Memory and CPU optimization (we optimize for M1max, 32Gb RAM)

Monitoring and Observability​

1. Logging (src/core/logging.py)​

Structured logging throughout the application

# Log levels:
- DEBUG     # Detailed debugging information
- INFO      # General application flow
- WARNING   # Potential issues
- ERROR     # Error conditions
- CRITICAL  # Critical failures

2. Health Checks​

# Health endpoints:
- /api/v1/health          # Basic health check
- /api/v1/status          # Detailed system status

Deployment Architecture​

1. Local Development Environment​

# docker-compose.yml
services:
  app:          # FastAPI application
  redis:        # Session storage
  qdrant:       # Vector database
  ollama:       # LLM service
  frontend:     # Web dashboard (optional)

Future Architecture Considerations​

1. Production Considerations​

  • Production optimization: multistage docker, with external cloud services
  • Cloud Database w Clustering: High availability setup
  • Monitoring Stack: Prometheus + Grafana
  • Log Aggregation: Centralized logging