Bootstrap Persona Architecture

Vision: Self-configuring AI that guides users from zero-config to optimal setup

🎯 Core Philosophy

"Install works immediately → AI demonstrates value → System suggests intelligent upgrades"

The Bootstrap Persona is a special AI that:

1. Awakens on first install (no configuration required)
2. Detects hardware capabilities (M1 MacBook Air baseline)
3. Tests available adapters (PEFT, Candle, cloud APIs)
4. Learns user preferences through conversation
5. Suggests optimal upgrades (MLX for Apple Silicon, DeepSeek for cloud)
6. Gracefully handles changes (API keys added/removed)

🏗️ Architecture Components

1. Hardware Detection Layer

interface HardwareProfile {
  platform: 'darwin' | 'linux' | 'win32';
  arch: 'arm64' | 'x64';
  memory: number;        // GB
  gpu: {
    type: 'apple-silicon' | 'nvidia' | 'amd' | 'none';
    model?: string;      // 'M1' | 'M2' | 'RTX 4090'
    vram?: number;       // GB
  };
  cpu: {
    cores: number;
    model: string;
  };
}

Detection Strategy:

• os.platform() → Platform
• os.arch() → Architecture
• os.cpus() → CPU info
• GPU detection via:
  • macOS: system_profiler SPDisplaysDataType (Metal/MPS)
  • Linux: nvidia-smi or rocm-smi
  • Fallback: No GPU detected

2. Adapter Health Check

interface AdapterStatus {
  adapterId: 'peft' | 'candle' | 'mlx' | 'deepseek' | 'openai';
  available: boolean;
  reason?: string;          // Why unavailable
  performance: {
    speed: 'fast' | 'medium' | 'slow';
    cost: number;           // $/training session
    requiresInternet: boolean;
    requiresAPIKey: boolean;
  };
  recommendation?: {
    priority: number;       // 1 = highest
    reason: string;
  };
}

Health Check Process:

1. PEFT: Check if Python environment bootstrapped
2. Candle: Check if Candle inference worker is available
3. MLX: Check if mlx importable (Apple Silicon only)
4. Cloud APIs: Check for environment variables

3. Bootstrap Persona Entity

interface BootstrapPersonaState {
  hardwareProfile: HardwareProfile;
  adapterStatuses: AdapterStatus[];
  userPreferences: {
    budget: 'free' | 'cheap' | 'performance';
    speed: 'patient' | 'balanced' | 'fast';
    privacy: 'local-only' | 'prefer-local' | 'cloud-ok';
  };
  recommendations: Recommendation[];
  checkpoints: {
    hardwareDetected: boolean;
    adaptersChecked: boolean;
    firstTrainingComplete: boolean;
    suggestedUpgrade: boolean;
  };
}

🔄 User Journey

Stage 1: Zero-Config Install

npm install
npm start

# Bootstrap Persona awakens
🤖 Bootstrap Persona: "Hi! I'm setting up Continuum on your M1 MacBook Air..."

[Detects hardware]
✅ Apple Silicon M1 detected
✅ 8GB unified memory
✅ MPS acceleration available

[Tests adapters]
✅ PEFT ready (local training)
✅ Candle inference ready
ℹ️  Cloud APIs not configured (optional)

🤖 "You're all set! Try chatting with the AI personas - they'll learn from you automatically."

Stage 2: Demonstration Phase

User chats with AIs, genome learning happens silently in background.

[After 10 interactions]
🤖 Bootstrap Persona: "I noticed you've had 10 conversations with Helper AI.
   Would you like me to train a personalized version? Takes ~30 seconds on your device."

[Yes] [Remind me later] [Tell me more]

Stage 3: Intelligent Upgrade Suggestions

[After 5 training sessions]
🤖 Bootstrap Persona: "I see you're training often! Here are some options:

   🚀 FASTER (Free):
      • Use Candle with GPU acceleration → 2x faster inference
      • Install MLX → 2x faster training (Apple Silicon native)

   💨 FASTEST (Paid):
      • DeepSeek API → Train in 5s instead of 30s
      • Cost: ~$0.0002 per session (extremely cheap)

   Your hardware: M1 MacBook Air (8GB) → MLX recommended

   [Install MLX] [Add API Key] [Keep Current] [Don't ask again]"

Stage 4: Graceful Degradation

[User removes API key]
🤖 Bootstrap Persona: "I noticed your DeepSeek API key was removed.

   ✅ No problem! Falling back to local PEFT training.
   ℹ️  Training will take ~30s instead of 5s, but still works perfectly.

   [Re-add key later] [OK]"

🧬 Genome Learning Integration

The Bootstrap Persona itself uses genome learning to:

1. Learn Hardware Patterns:

   • Track training times for different adapters
   • Measure actual performance vs estimates
   • Detect when system is under load

2. Learn User Preferences:

   • How often user trains
   • Response to upgrade prompts
   • Budget sensitivity signals

3. Improve Recommendations:

   • Fine-tune suggestion timing
   • Personalize message tone
   • Optimize cost/performance balance

Checkpoint Strategy

// Stored in .continuum/genome/bootstrap-persona-checkpoint.json
{
  "version": "1.0",
  "hardwareProfile": { /* detected on install */ },
  "adapterHistory": [
    {
      "adapterId": "peft",
      "trainingSessions": 10,
      "avgTrainingTime": 28.3,
      "lastUsed": "2025-11-02T19:00:00Z"
    }
  ],
  "userInteractions": {
    "upgradePromptShown": true,
    "upgradePromptResponse": "remind-later",
    "trainingsCompleted": 10,
    "preferredAdapter": "peft"
  },
  "genomeState": {
    "adapterLoaded": null,  // No adapter initially
    "traitsLearned": [
      "hardware-detection",
      "cost-awareness",
      "timing-optimization"
    ]
  }
}

📊 Recommendation Algorithm

function recommendAdapters(
  hardware: HardwareProfile,
  preferences: UserPreferences,
  history: TrainingHistory
): Recommendation[] {
  const recommendations: Recommendation[] = [];

  // Always ensure local fallback
  recommendations.push({
    adapterId: 'peft',
    priority: 1,
    reason: 'Universal fallback - works on any hardware',
    action: 'keep-enabled'
  });

  // Apple Silicon optimizations
  if (hardware.gpu.type === 'apple-silicon') {
    if (!hasAdapter('mlx')) {
      recommendations.push({
        adapterId: 'mlx',
        priority: 2,
        reason: `2x faster training on ${hardware.gpu.model}`,
        action: 'suggest-install',
        estimatedImprovement: '50% faster',
        installCmd: 'pip install mlx mlx-lm'
      });
    }

    if (!hasAdapter('candle')) {
      recommendations.push({
        adapterId: 'candle',
        priority: 3,
        reason: 'Faster inference with Metal acceleration via Candle',
        action: 'suggest-install'
      });
    }
  }

  // Cloud recommendations (budget-aware)
  if (preferences.budget !== 'free' && history.trainingSessions > 5) {
    const avgTime = history.avgTrainingTime;
    const cloudTime = 5; // ~5s with API
    const savings = avgTime - cloudTime;

    if (savings > 20) { // Worthwhile if saves 20+ seconds
      recommendations.push({
        adapterId: 'deepseek',
        priority: 4,
        reason: `Train in ${cloudTime}s instead of ${avgTime}s`,
        action: 'suggest-api-key',
        estimatedCost: 0.0002,
        estimatedImprovement: `${Math.round(savings)}s faster`
      });
    }
  }

  return recommendations.sort((a, b) => a.priority - b.priority);
}

🎨 UI/UX Patterns

Timing Strategy

DON'T show upgrade prompts:

• On first install (too early)
• During active chat (disruptive)
• More than once per day (annoying)

DO show upgrade prompts:

• After 5-10 training sessions (proven value)
• When training times exceed threshold (pain point visible)
• After user expresses interest in speed
• When system detects underutilized hardware

Message Tone Examples

Informative, not pushy:

✅ "Your training completed in 28s - perfectly normal for M1!"
💡 "Tip: Candle with GPU acceleration could reduce this to 15s (free, local)"
[Tell me more] [Maybe later]

Celebratory, not salesy:

🎉 "Nice! You've trained 10 times. Your AIs are getting smarter!"
🚀 "Want even faster results? I have some ideas..."
[Show me] [I'm happy with current speed]

Empowering, not technical:

🧠 "I noticed your M1 has Metal acceleration"
✨ "This means you can use MLX for 2x faster training (free!)"
📦 "One command: pip install mlx"
[Install now] [Explain more] [Not interested]

🔐 Privacy & Control

The Bootstrap Persona:

• Never sends hardware data externally (all local)
• Asks permission before installing anything
• Respects "don't ask again" choices
• Allows downgrading (remove adapters/keys anytime)
• Documents all recommendations (explainable AI)

User controls:

# Disable Bootstrap Persona suggestions
./jtag config/set --key="bootstrap.suggestions.enabled" --value=false

# Reset Bootstrap Persona state
./jtag genome/reset --persona=bootstrap

# View current recommendations
./jtag genome/recommendations

🚀 Implementation Phases

Phase 1: Foundation (Current) ✅

• PEFT adapter working
• Python environment bootstrapped
• Integration tests passing

Phase 2: Bootstrap Persona Core (Next)

• Hardware detection utility
• Adapter health checks
• Bootstrap Persona entity creation
• Checkpoint system

Phase 3: Intelligence Layer

• Recommendation algorithm
• Timing optimization
• Message generation
• UI integration

Phase 4: Adapter Expansion

• Candle GPU acceleration (hybrid mode)
• MLX adapter (Apple Silicon native)
• Cloud API adapters (DeepSeek, OpenAI)

Phase 5: Self-Healing

• Automatic fallback on adapter failure
• API key rotation support
• Performance regression detection
• Proactive troubleshooting

🎯 Success Metrics

The Bootstrap Persona is successful when:

1. Zero-config works: 90%+ users complete first training without issues
2. Discovery rate: 50%+ users discover upgrade options naturally
3. Adoption rate: 30%+ users adopt at least one suggested upgrade
4. Satisfaction: 80%+ users find recommendations helpful
5. Retention: Users who adopt upgrades are 2x more likely to continue using system

🔮 Future Vision

Self-Optimizing System:

• Bootstrap Persona trains itself on thousands of hardware profiles
• Learns optimal adapter combinations for each device type
• Shares anonymized learnings across installations (optional)
• Becomes expert at matching hardware to workloads

Community Intelligence:

• "1000 M1 users found MLX 2.3x faster than PEFT"
• "DeepSeek most cost-effective for small training runs (<100 examples)"
• "Candle + PEFT hybrid gives best balance on M1"

Adaptive Personas:

• Bootstrap Persona evolves into "Setup Concierge" after onboarding
• Later helps with advanced features (multi-model, distributed training)
• Eventually trains replacement (user's custom setup advisor)

---

Philosophy: Every user's hardware is different, but the path to genomic AI should be universal - automatic, intelligent, and always improving.