AegisDominus: Secure AI Assistant with Full Observability

Overview

AegisDominus is a security-first AI proxy designed to protect Large Language Models (LLMs) from malicious misuse. It sits between the user and Google Gemini, providing real-time threat detection and granular observability that standard API calls lack.

Key Capabilities:

1. Active Defense: Automatically detects and blocks "Jailbreak" attempts (e.g., prompt injection attacks) before they process.
2. Deep Observability: Uses Datadog APM to trace every interaction, measuring token consumption, latency, and error rates per user.
3. Reliability Engineering: Implements Service Level Objectives (SLOs) to guarantee 99% system uptime and performance.
4. Email: Automated alert email pushed to your email id.

How it works

• Core Backend: Python (Flask)
• AI Engine: Google Vertex AI (gemini-2.0-flash-001)
• Observability: Datadog APM & Tracing

---

Quick Start Guide

Prerequisites

1. Google Cloud Project with Vertex AI API enabled.
2. Datadog Account (and API Key).
3. Google Cloud CLI (gcloud) installed and authenticated. (its easier to download the exe file to install gcloud CLI).

Step 1: Clone & Install

git clone https://github.com/PS-003R32/AegisDominus.git
cd AegisDominus

# Install dependencies
pip install -r requirements.txt

Step 2: Google Cloud Authentication

Authenticate using your local Google credentials to access the Gemini model:

# Login to Google Cloud via gcloud CLI
gcloud auth application-default login

# Set your Project ID (Required for Vertex AI)
export PROJECT_ID="your-google-project-id" #create one and check in the settings
export LOCATION="us-central1"

Step 3: Configure Datadog

Export your Datadog credentials directly in the terminal:

export DD_API_KEY="your_datadog_api_key_here"
export DD_SITE="us5.datadoghq.com"  # Check your account (e.g., datadoghq.com)
export DD_ENV="devpost-demo" # optional
export DD_SERVICE="aegis-dominus" # optional

Step 4: Run the Application

Start the server using the Datadog Tracer wrapper to enable observability.
[NOTE: I have used python venv to run this project. YOu may proceed without setting it up.]

ddtrace-run python src/app.py

Testing & Simulation

To demonstrate the security features, run the included traffic generator. This script simulates valid users and malicious attackers.

python traffic_generator.py

---

Dashboard & Observability

The AegisDominus Command Center (Datadog Dashboard) visualizes the system's health and security status with 5 key widgets: [NOTE: You can find and download the datadog dashboard jason file in this repository to make it easier.]

1. Total Tokens Generated: Tracks AI consumption costs (Query Value).
2. Security Incidents: A real-time graph that spikes RED when http.status_code:403 (Blocked Attacks) occur.
3. Traffic Volume: Monitors the heartbeat of total requests hitting the system.
4. Security Rules: A "Monitor Summary" list that instantly flips from OK (Green) to ALERT (Red) during an attack.
5. SLO Reliability: A live badge tracking our 99% uptime target over a 7-day window.
   

[Once you setup the simulation you will recieve email alerts sent automatically through Datadog]

Terminal view

Live Datadog Dashboard

Email Alerts

---

LICENSE

This project is licensed under the MIT License – see the LICENSE file for details.