"Every existing system tells you what happened. ClawSentinel tells you if it matters. It is the first system that knows the difference between your cat and a criminal."
DEMO LINK: https://youtu.be/_dhuv23u6BU
PRESENTATION LINK: https://canva.link/fzn91jknohnwxxy
The ClawSentinel 3D Command-and-Control Interface triggering a Telegram contextual alert.
- The Problem
- The Solution
- Three-Layer AI Architecture
- Key Features
- Real-World Use Cases
- Tech Stack
- Impact
- The Moat
- Getting Started
- Hackathon Context
Modern smart homes are noisy. Current systems (Ring, ADT, basic SmartThings) rely on fixed rules that lack nuance.
- Zero Context: They trigger on everything—pets, shadows, or normal evening routines.
- Alert Fatigue: ~73% of users eventually mute or ignore notifications due to frequent false alarms.
- Threat Blindness: Because the system is muted or ignored, real emergencies are lost in the sea of trivial data.
ClawSentinel transforms reactive hardware into a proactive guardian. By adding a behavioral "soul" to standard sensor data, it filters out the noise of daily life and only interrupts the user when a high-risk anomaly occurs.
- Context-Aware: Knows the difference between your cat and a criminal.
- Behavioral Memory: Uses a memory-first architecture to understand what is "normal" for your home.
- Explainable AI: Don't just get an alert; know exactly why the system flagged it.
ClawSentinel acts as a sophisticated intelligence layer on top of hardware ecosystems like Samsung SmartThings.
graph LR
A[SmartThings Sensors] --> B[SmartThings API]
B --> C[ClawSentinel AI Engine]
subgraph AI Engine
C --> D[ML Anomaly Detection]
D --> E[OpenClaw Contextual Reasoning]
E --> F[Behavioral Memory]
end
F --> G[Telegram Smart Alert]
G --> H[User Action / Auto-Response]
H --> I[SmartThings Actions]
High-performance FastAPI gateway handling telemetry. This layer normalizes disparate data into a unified event stream for analysis.
Utilizes Scikit-Learn Isolation Forests trained on 90 days of historical data.
- Pattern Learning: Recognizes that hallway motion at 2 PM is normal, but a back door opening at 3 AM is a high-risk anomaly.
- 99% Noise Reduction: Filters out mundane activity instantly, ensuring the heavier reasoning layers only activate for true threats.
The "Brain" of the system. An orchestration of specialized agents that reason over context.
- Stateful Memory: Uses
SOUL.mdandHEARTBEAT.mdto maintain long-term behavioral context. - Reasoning: "The back door opened, but the owner's phone is connected to the home Wi-Fi — this is a safe return, not an intrusion."
- Multi-Agent Orchestration: Powered by OpenClaw to coordinate Sensor, Risk, Decision, and Action agents.
- Spatial 3D Dashboard: Immersive threat visualization built with React Three Fiber (Three.js).
- Local-First Privacy: Core intelligence runs locally to prevent IoT breach vulnerabilities.
- Interactive Control: Secure two-way command and control via Telegram Bot API.
Our anomaly detection model trains on historical data to build a highly accurate behavioral baseline, drastically reducing false positives.
Hourly activity heatmap showing the model clearly learning normal sensor windows, with darker bars indicating higher volume.
Weekly event patterns demonstrating tight weekday consistency versus more variable weekend behavior.
Baseline learning in action: As ClawSentinel learns the home's routines, false alert rates drop sharply over time.
| Scenario | Trigger | Action |
|---|---|---|
| The Intrusion | 3 AM motion, user confirmed away. | High-Risk Flag. Pings Telegram: "Lock door & alert security?" |
| The Mid-Day Delivery | Front door activity at 2 PM. | Suspicious, not Dangerous. Matches typical window. Logs event silently. |
| The Nightly Pet | Hallway motion at 1 AM. | Normal. Recognizes household pet baseline. Zero false alarm. |
| The Safe Return | Door opens, user returns home. | Authorized Entry. Recognizes routine. Silently disarms and welcomes user. |
Recognizing a routine homecoming and adjusting security state without user intervention.
Analyzing timing and occupancy to differentiate a courier from a threat.
Filtering low-height movement baselines to eliminate false positives.
- Google Gemini 1.5 Flash: High-level contextual reasoning for the Decision Agent.
- River & Scikit-Learn: Real-time, adaptive anomaly scoring (Isolation Forest) on live streaming sensor data.
- OpenClaw: The central nervous system orchestrating all specialized agents.
- Stateful Memory: File-based architecture for persistent behavioral context preservation.
- FastAPI: Asynchronous, production-ready ASGI engine with rigorous validation.
- Telegram Bot API: Secure, real-time interactive command and control.
- React + Three.js: Spatial environment mapping for 3D threat visualization.
- Vite + Zustand + Tailwind: Ultra-fast builds, reactive state, and modern glassmorphism UI.
- ZERO Alert Fatigue: Eliminates the noise that leads users to disable their security systems.
- 90-Day Intelligence Baseline: "Zero cold-start" AI that dynamically learns your home's unique patterns.
- 99.9% Local Privacy: Core intelligence processed locally for maximum security.
- Enterprise Grade: Brings sophisticated behavioral reasoning to standard OTC hardware.
- Behavioral Intelligence: Hard to replicate as it requires personalized data and site-specific modeling.
- On-Device AI: Combines ML + Persistent Memory + LLM Reasoning locally.
- Compound Learning: The system gets smarter every single day it lives in your environment.
Follow these steps to boot the entire ClawSentinel ecosystem locally.
cd backend
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env # Add your GEMINI_API_KEY and TELEGRAM_TOKEN
python main.py# In a new terminal
cd frontend
npm install
npm run dev# In a third terminal to stream mock sensor data
cd ml
python stream_model.pyThis project was developed for the Samsung PRISM program to showcase the potential of AI-driven intelligence layers on top of existing smart home hardware ecosystems.
- Member 1: Sreeya Chand
- Member 2: Prapti
- Member 3: Aniksha Anithan
- Member 4: Samyukthaa M
ClawSentinel — Giving smart environments a behavioral soul.