• 

  NeonClaw Workflow and the design

• 

  the backened design

• 

NemoClaw: Distributed Sovereign Personal AI

Elevator Pitch (Devpost Short Tagline)

A dual personal AI agent running on laptop and home server, powered by NVIDIA Nemotron, OpenShell, Hermes Agent, and Nebius Serverless with sovereign memory and autonomous workflows.

💡 Inspiration

Cloud-based AI assistants require users to sacrifice data privacy, lock workflows into centralized silos, and rely on rigid, rule-based responses.

I wanted a personal AI agent that runs across my home server and laptop as one unified mind—an always-on sovereign system that can intelligently offload workloads between devices. When I'm away, my home server handles 24/7 background email triage, infrastructure monitoring, and stealth web research. When I sit at my workstation, it coordinates directly with my laptop edge worker to execute local git audits, run tests, and inspect code—all without sending a single byte of private data to third-party proprietary clouds.

🚀 What NemoClaw Does

NemoClaw is an autonomous, scattered-agent personal AI system orchestrated across a private Home Server mesh (homelabb) and a Laptop Edge Workstation (nemoclaw-edge):

• ⚡ Concurrent Dual-Node Parallel Execution: Dispatches tasks simultaneously across both nodes (e.g., home server monitors live Docker container health while the laptop workstation audits git repositories and verifies tests) returning a synthesized report in under 2 seconds.
• 📬 Autonomous Email Urgency Triage: Securely connects via IMAP/SMTP to triage incoming messages with intelligent heuristic urgency classification (🚨 URGENT outages & security incidents, ⚠️ HIGH client deadlines & proposals, 💤 LOW promotional digests) and autonomously drafts contextual replies.
• ☕ Executive Morning Briefing (/briefing): Compiles a scheduled daily intelligence digest combining inbox email urgency, live dual-device hardware health (server CPU/RAM + laptop CPU/RAM/battery), and AI research headlines.
• 🌐 Obscura Stealth Headless Browser: Autonomous web research, site scraping, and screenshot capture with authenticated session isolation.
• 🔒 Dual-Tier Sovereign Memory Vaults: Cryptographic separation between Shared Memory (reusable skills, user preferences) and Personal Vaults (IDOR-protected private telemetry and local credentials).
• 🎙️ Telegram Interface (Voice & Text): A conversational chief-of-staff experience with Whisper STT voice note processing and executive concierge travel & booking coordination.

⚙️ How We Built It & NVIDIA / Nebius Resources Used

We anchored the architecture on NVIDIA's state-of-the-art open models and Nebius Serverless cloud acceleration:

• Brain Engine — NVIDIA Nemotron 70B (nvidia/llama-3.1-nemotron-70b-instruct): Drives all agentic reasoning, multi-turn conversational context, tool-call parameter extraction, and contextual synthesis.
• Nebius Serverless Studio: Powers high-throughput, ultra-low-latency API inference for the Nemotron foundation model with automatic fallback routing.
• NVIDIA NemoClaw & OpenShell Sandboxing: Enforces strict container isolation, secure command sandboxing, and policy-driven execution across edge devices.
• Hermes Agent Cognitive Loop: Implements multi-step tool execution where tool observations are fed back into Nemotron for structured second-pass synthesis.
• Scattered Mesh Architecture:
  • Home Server (homelabb): Runs 8 orchestrated microservices in Docker (neonmesh, neoncore, neonvault, neonmem, obscura, neonshell, neonpulse, redis).
  • Laptop Edge Worker (nemoclaw-edge): Lightweight Python daemon reporting real-time CPU/RAM/battery telemetry and executing workstation-local tasks over an encrypted Tailscale P2P tunnel.
  • Obscura Engine: Playwright headless browser for stealth web extraction.

⚡ Challenges We Ran Into

• Eliminating Robotic Fallbacks: Chatbots often sound mechanical ("I received your message"). We completely re-engineered the prompt architecture and fallback reasoners to maintain an executive Jarvis/chief-of-staff persona with real-time concierge actions.
• True Parallel Multi-Node Orchestration: Coordinating tasks across physically separate machines without latency bottlenecks required an asynchronous non-blocking event loop using asyncio.gather and synchronized SQLite state.
• Accurate Urgency Scoring: Building email triage that avoids alert fatigue by distinguishing genuine production downtime from promotional noise using multi-keyword scoring.

🏆 Accomplishments We're Proud Of

• 10/10 Test Suites Passing: 100% test pass covering dual-vault IDOR security, Hermes multi-tool reasoning, email triage, voice transcription, and parallel dual-node execution.
• Sub-2-Second Dual-Node Execution: Concurrent task offloading across server and laptop simultaneously.
• Zero Data Leakage: All memories, vaults, and credentials remain private on our sovereign hardware.

🔮 What's Next for NemoClaw

• Local Nemotron Quantizations: Running quantized Nemotron models directly on laptop TensorRT-LLM for offline flight/travel mode.
• Autonomous Calendar Negotiation: Meeting scheduling and multi-agent coordination with other sovereign NemoClaw nodes.

🛠️ Built With

nvidia-nemotron nebius-serverless hermes-agent openshell python fastapi docker redis nodejs sqlite playwright tailscale telegram-bot-api

Built With

• hermes-agent
• javascript
• nvidia
• openshell
• python
• sql