Maximus-X Sentinel 🧠

A private, GPU-accelerated personal AI brain. Nobody else has built this.

A fully local, multi-agent personal AI assistant running on NVIDIA Jetson Orin Nano with Raspberry Pi 5 voice edge and Mac dashboard. Combines OpenClaw multi-channel messaging gateway, LangGraph agent orchestration, Qdrant vector memory, and Ollama inference — all on-premise, zero cloud, zero subscription.

What Makes This Different

Most "personal AI" setups are either:

• Cloud-dependent (ChatGPT, Claude.ai, etc.) — your data leaves your machine
• Single-model chatbots — no specialization, no memory, no autonomy

Maximus-X Sentinel is different:

Feature	What it does
OpenClaw Gateway	Talk to Maximus via WhatsApp, Telegram, Signal, Discord, iMessage — all routed to your Jetson
LangGraph Supervisor	Intelligent routing to specialized sub-agents (Research, ChemBiz, Home, Schedule)
Context Membrane	Auto-ingesting RAG layer that pulls from your local notes, emails, and docs nightly
Jetson Inference Core	GPU-accelerated local LLM — no API keys, no token costs, no privacy leaks
Pi 5 Voice Edge	Wake word + faster-whisper STT + Kokoro TTS, runs on Pi hardware
Self-improving	OpenClaw can write and install its own new skills when you ask for new capabilities

---

System Architecture

┌─────────────────────────────────────────────────────────────┐
│                     YOUR DEVICES                            │
│  WhatsApp / Telegram / Signal / iMessage / Discord          │
└────────────────────────┬────────────────────────────────────┘
                         │
┌────────────────────────▼────────────────────────────────────┐
│              Mac / Laptop  (OpenClaw Gateway)               │
│  • openclaw gateway process (Node.js)                       │
│  • Open WebUI dashboard  :3000                              │
│  • Routes all channels → Jetson agent                       │
└────────────────────────┬────────────────────────────────────┘
                         │ LAN / Wi-Fi
┌────────────────────────▼────────────────────────────────────┐
│           NVIDIA Jetson Orin Nano  (AI Brain)               │
│                                                             │
│  ┌─────────────────────────────────────────────────┐        │
│  │         LangGraph Supervisor Agent              │        │
│  │   Routes to: Research | ChemBiz | Home | Sched  │        │
│  └──────────┬──────────────────────┬──────────────┘        │
│             │                      │                        │
│  ┌──────────▼──────┐    ┌──────────▼──────┐                │
│  │  Ollama (LLM)   │    │  Qdrant (RAG)   │                │
│  │  llama3.2:3b    │    │  Context Membrane│               │
│  │  GPU accel.     │    │  Auto-ingestion  │                │
│  └─────────────────┘    └─────────────────┘                │
│                                                             │
│  FastAPI server :8000 ← all agent traffic                   │
└────────────────────────┬────────────────────────────────────┘
                         │
┌────────────────────────▼────────────────────────────────────┐
│             Raspberry Pi 5  (Voice Edge)                    │
│  • openWakeWord  (wake: "Hey Maximus")                      │
│  • faster-whisper STT (CTranslate2 backend)                 │
│  • Kokoro-82M TTS (HF TTS Arena #1, offline)                │
│  • Home/work presence triggers                              │
└─────────────────────────────────────────────────────────────┘

---

Prerequisites

• Jetson Orin Nano running JetPack 6.4+
• NVMe SSD strongly recommended (SD card I/O is the bottleneck for model loading)
• Raspberry Pi 5 (4GB or 8GB)
• Mac/Linux laptop for OpenClaw gateway
• Ollama model pulled: ollama pull llama3.2:3b

---

Quick Start

1. Clone & configure

git clone https://github.com/shehanmakani/MaximusX
cd MaximusX
cp .env.example .env
# Edit .env — set OPENCLAW_TOKEN, TELEGRAM_BOT_TOKEN, etc.

2. Start the Jetson stack

docker compose up -d

3. Start OpenClaw on Mac

npm install -g @openclaw/openclaw
openclaw init
openclaw start

4. Start voice edge on Pi

cd pi-voice
pip install -r requirements.txt
python3 voice_edge.py

Autonomous twin starter

This repo now includes a local self-prompting loop that uses future-self-emulator to rank likely next actions from calendar and repo context, then stages the winning task for approval instead of executing it blindly.

python3 self_prompting_agent.py \
  --profile "Shehan: founder, engineer, builder" \
  --calendar sample_calendar.json \
  --repo .

That writes a staged task into autonomous_outputs/task_<timestamp>.json.

Approve a staged task locally:

python3 approve_task.py <task_id>

Or through the WhatsApp gateway:

• GO → generate the latest predicted task
• STATUS → show the latest pending task
• APPROVE <task_id> → execute the allowlisted action

Sleep mode workflow

This repo also includes a first pass at the overnight clone workflow:

1. Log daytime activity and repo state:

python3 activity_logger.py --repo .
python3 activity_logger.py --event-type meeting_prep --summary "Prepared architecture notes before the IntelliForm sync"

2. Enter sleep mode before bed:

python3 sleep_mode.py \
  --profile "Shehan: founder, engineer, builder" \
  --calendar sample_calendar.json \
  --repo . \
  --cycles 2

3. Wake up and review the digest:

python3 wake_mode.py
python3 morning_digest.py --json

Core files:

• activity_logger.py captures daytime behavior into memory/activity_log.jsonl
• pattern_model.py learns recurring priorities from those events
• reasoning_loop.py critiques and ranks options using strategic, practical, protective, and identity lenses
• overnight_planner.py converts those patterns into a ranked overnight task queue
• sleep_mode.py stages tasks while you are "asleep"
• wake_mode.py gives you the morning approval summary

Important design note: This system should not just copy previous behavior. The newer overnight flow now includes a reasoning loop so it can adapt when priorities change, critique its own options, and prefer work that matches both current context and the user's style rather than blindly repeating patterns.

---

Agent Specializations

Agent	Trigger keywords	Capabilities
Research	"find", "search", "what is", "summarize"	Web search, RAG over your docs
ChemBiz	"chemrich", "intelliform", "formulation", "lead"	Chemical domain Q&A, business context
Home	"lights", "temperature", "lock", "scene"	Home Assistant REST API
Schedule	"remind", "meeting", "calendar", "when"	Google Calendar + cron reminders

---

OpenClaw Skills Included

• chembiz-context — Loads ChemRich/ChemeNova domain knowledge into every ChemBiz query
• nightly-ingest — Cron job: pulls new emails/notes into Qdrant Context Membrane at 2am
• voice-relay — Bridges Pi STT output → OpenClaw → Jetson and back to Pi TTS
• presence-trigger — Detects home/work arrival via Pi sensor, fires contextual briefing

---

Tech Stack

Layer	Technology
LLM inference	Ollama + llama3.2:3b (Jetson GPU)
Agent orchestration	LangGraph + LangChain-core
Vector DB	Qdrant v1.13.0 (arm64)
Messaging gateway	OpenClaw (self-hosted)
Dashboard	Open WebUI
Voice STT	faster-whisper (CTranslate2)
Wake word	openWakeWord
TTS	Kokoro-82M
Container runtime	Docker + NVIDIA runtime
API server	FastAPI + Uvicorn

---

SSH Key Exchange: Ensure your Pi 5 can SSH into the Jetson without a password: ssh-copy-id shehan@192.168.1.100

Ngrok Setup: On the Pi, run ngrok http 5000. Copy the URL into the Twilio Console under "Sandbox Settings" > "When a message comes in."

Privacy Guarantee

• Zero cloud inference — all LLM calls stay on Jetson

• OpenClaw stores config/memory as local Markdown on your Mac

• Qdrant vector data stays on Jetson NVMe

• Only outbound: your chosen messaging app (Telegram, etc.) for delivery

  ---

🔮 Also Check Out

If you're interested in decision intelligence, check out Future-Self-Emulator — an AI system that simulates multiple life-path versions of yourself using Agentic AI.