SentinAI Whitepaper

Autonomous Operations for Layer 2 Rollup Infrastructure

Version 1.0 | February 2026

---

Executive Summary

As Layer 2 rollup infrastructure grows in complexity, manual operational oversight becomes increasingly untenable. SentinAI is an autonomous node guardian designed specifically for Optimism-based rollup infrastructure, combining real-time telemetry, AI-powered anomaly detection, and policy-governed execution to detect, diagnose, and remediate operational issues with minimal human intervention.

Unlike black-box autopilots, SentinAI implements a safety-first autonomy model: low-risk actions execute automatically, high-risk operations require explicit approval, and every decision is auditable.

---

Key Highlights

🎯 Problem We Solve

Modern L2 rollup deployments face:

• Operational complexity: Multiple interdependent components (op-geth, op-node, op-batcher, op-proposer)
• High MTTR: 30-60 minutes average response time with manual operations
• Inconsistent execution: Remediation quality varies by operator experience
• 24/7 burden: Service degradation during off-hours

🛡️ Our Approach: Governed Autonomy

Safety-First Design:

• Hard-coded blacklist prevents destructive actions
• Risk-tiered execution (Low/Medium/High/Critical)
• Default dry-run mode for testing

Policy-Over-Model Execution:

• Explainable decision trees augmented by AI
• No opaque black-box models
• Graceful degradation when AI unavailable

Auditability by Default:

• Every decision logged with reasoning
• Exportable audit trails for compliance
• Post-mortem analysis support

🏗️ System Architecture

Six Core Subsystems:

1. Telemetry Collector: Aggregate metrics from L2 RPC, Kubernetes API, component logs
2. Anomaly Detection: Statistical (z-score) + AI log analysis (Claude Haiku 4.5)
3. Root Cause Analysis: AI-guided diagnosis with human-readable reasoning
4. Predictive Scaling: Time-series forecasting for proactive resource allocation
5. Action Planning & Execution: Risk-based policy framework with automatic rollback
6. MCP Integration: External AI agent access (Claude Desktop, Claude Code)

System Data Flow:

graph TB
    A[L2 RPC<br/>eth_blockNumber<br/>eth_gasPrice<br/>txpool_status] -->|30s polling| B[Ring Buffer<br/>60 points<br/>30-min window]
    B --> C[Anomaly Detection<br/>z-score > 2.0]
    C -->|Anomaly Event| D[RCA Engine<br/>Claude Haiku 4.5]
    D -->|Root Cause| E[Action Planning<br/>Policy-based]
    E -->|Low/Medium Risk| F[Auto-Execute]
    E -->|High/Critical Risk| G[Approval Gate]
    F --> H[Verification]
    G -->|Approved| F
    H -->|Success| I[Complete]
    H -->|Failure| J[Automatic Rollback]
    
    style A fill:#e1f5ff
    style C fill:#fff3cd
    style D fill:#f8d7da
    style E fill:#d1ecf1
    style F fill:#d4edda
    style G fill:#f8d7da
    style H fill:#fff3cd
    style I fill:#d4edda
    style J fill:#f8d7da

🔒 Risk & Control Framework

Forbidden Actions (Hard-coded):

• DROP DATABASE
• Table-wide DELETE
• kubectl delete namespace/serviceaccount
• kubectl exec without approval

📈 Case Studies

1. Sync Stall Recovery

• Incident: op-node fell 50 blocks behind L1
• Detection: blockInterval anomaly (4.2s → 12.8s, z=4.1)
• Action: Auto-restart op-node
• Result: MTTR 3.4 min vs. 45 min baseline ✅

2. Batcher Congestion

• Incident: L1 gas spike delayed batch submissions
• Detection: txPoolCount anomaly (23 → 187, z=3.8)
• Action: Increase gas budget (approval required)
• Result: MTTR 12 min vs. 60 min baseline ✅

3. CPU Pressure Scaling

• Incident: CPU spike to 89% during traffic surge
• Detection: cpuUsage anomaly (45% → 89%, z=3.5)
• Action: Auto-scale to 4 vCPU
• Result: MTTR 2.8 min, prevented degradation ✅

---

Full Whitepaper (Academic Version)

For the complete technical whitepaper with detailed architecture, evaluation methodology, security analysis, and roadmap:

Sections in Full Whitepaper:

1. Problem Statement (Operational complexity, manual limits, autopilot dilemma)
2. Design Principles (Safety-first, policy-over-model, auditability)
3. System Architecture (6 subsystems with detailed flow)
4. Incident Lifecycle (Detect → Plan → Approve → Verify → Rollback)
5. Risk & Control Framework (Tiers, blacklist, approval boundaries)
6. Evaluation Metrics (MTTR, auto-resolution rate, false action rate)
7. Case Studies (3 real-world scenarios)
8. Security & Compliance (Least privilege, traceability, audit controls)
9. Roadmap (Q1/Q2 2026, future research)
10. Limitations & Future Work
11. Conclusion & Adoption Path

---

Roadmap

Q1 2026 (Current)

• ✅ Core autonomy engine with risk tiers
• ✅ MCP integration for external agents
• 🚧 Multi-cluster support
• 🚧 Prometheus/Grafana integration

Q2 2026

• Self-healing feedback loop
• Cost optimization engine
• Multi-model ensemble predictions
• Webhook notifications (Slack, Discord, PagerDuty)

Future Research

• Causal inference for root cause graphs
• Adversarial testing (chaos engineering)
• Cross-chain coordination

---

Learn More

Documentation & Resources

• Documentation: https://sentinai-xi.vercel.app/docs
• GitHub Repository: https://github.com/tokamak-network/SentinAI
• Quick Start Guide: 5-minute setup guide
• Architecture Deep Dive: System design documentation
• API Reference: Complete API documentation

Community & Support

• Contact: contact@sentinai.ai
• Issues & Feature Requests: GitHub Issues

---

Acknowledgments: This work builds on open-source contributions from the Optimism, Ethereum, and AI research communities.