Sentinel Dual-Engine Risk Framework

Sentinel is a high-performance risk assessment platform that allows real-time switching between Cloud-scale inference (Cerebras) and Private Local execution (NVIDIA GPU).

1. Frontend Development Setup

To run the Sentinel UI locally on your machine:

1. Prerequisites: Ensure you have Node.js installed (which includes npm).
2. Initialize Project:

   # Create a new directory and enter it
   mkdir sentinel-app && cd sentinel-app
   
   # Initialize a Vite project (React + TypeScript)
   npm create vite@latest . -- --template react-ts
   
   # Install dependencies used in this app
   npm install lucide-react recharts @google/genai

3. Copy Files: Place the App.tsx, index.html, types.ts, and the components/ & services/ folders into your local project structure.
4. Run Development Server:

   npm run dev

5. Access the App: Open your browser to http://localhost:5173.

---

2. Local NVIDIA Setup (Ollama)

To utilize the NVIDIA Local Mode, you must have a local LLM server running on your machine using Ollama.

Setup Steps

1. Install Ollama: Download and install from ollama.com.
2. Pull a Risk Model: Open your terminal and pull a high-reasoning model:

   ollama pull llama3

3. Ensure API Accessibility: Ollama serves an OpenAI-compatible API by default on http://localhost:11434/v1.
4. Hardware Optimization: Ensure your NVIDIA GPU (RTX 3090/4090 recommended) is correctly recognized by Ollama. You can check this by running ollama run llama3 and monitoring your GPU usage in Task Manager or nvidia-smi.

---

3. Architecture Overview

• Cloud Engine (Cerebras): Routes requests via the Cerebras SDK for sub-second inference on Wafer-Scale Engines. Best for high-throughput, non-private data.
• Local Engine (NVIDIA/Ollama): Routes requests via the local loopback to your GPU. Best for sensitive data privacy, offline testing, and cost-free execution.

4. Hardware Requirements

• NVIDIA RTX GPU: Pascal architecture or newer (Ampere/Ada Lovelace preferred for FP16/BF16 performance).
• VRAM:
  • 8GB for Llama3 8B (Quantized)
  • 24GB for Llama3 70B (Quantized) or larger models.
• Drivers: Latest NVIDIA Game Ready or Studio drivers.

5. Integration Details

The system tracks and compares:

1. Network Latency: Time for packets to reach Cerebras Cloud.
2. VRAM Latency: Time for the local GPU to process weights and generate tokens via Ollama.