AegisAI

Helping Students Build Safer AI, Faster

Comment

🌸 AegisAI — Helping Students Build Safer AI, Faster

Where AI ideas bloom without breaking.

🎯 Inspiration

The inspiration for AegisAI started with a simple problem many students and beginner developers face:

“My AI app works… but how do I know if it’s safe, efficient, or silently failing?”

As students build chatbots, AI tools, and LLM-powered applications, they quickly encounter invisible issues:

  • A small prompt change suddenly explodes API costs
  • A harmless-looking input causes unsafe or unexpected AI behavior
  • A model hallucination exposes sensitive information
  • The app works, but no one knows what’s happening under the hood

Unlike traditional software, LLMs are probabilistic black boxes.
There are no stack traces for hallucinations, no debuggers for prompt injection, and no warnings when token usage quietly drains budgets.

Most existing observability and security tools are built for experts and enterprises, not students.

This led to one guiding question:

What if students had a simple safety net while building AI apps?

Something that watches AI behavior in real time, explains failures in plain English, and helps beginners fix issues instead of guessing.

That idea became AegisAI.

💡 What AegisAI Does

AegisAI is an AI safety and observability assistant designed for students, beginners, and indie developers building LLM-powered applications.

Think of it as:

A black box flight recorder for AI apps — built for learning.

AegisAI wraps around an AI application and continuously monitors:

  • Security risks
  • Performance issues
  • Token usage and cost anomalies
  • Unexpected or unsafe AI behavior

Instead of overwhelming users with raw logs, AegisAI explains what went wrong, why it matters, and how to fix it.

✨ Core Features

🚨 Real-Time AI Safety Detection

  • Detects 16 common prompt injection and misuse patterns
  • Covers instruction override, role abuse, and data leakage attempts
  • Assigns clear severity levels (Low → Critical)
  • Designed to help beginners understand why something is risky

🧠 AI-Powered Explanations

  • Plain-English “AI autopsy reports”
  • Explains:
    • What happened
    • Why it matters
    • How to fix it
  • Side-by-side prompt improvement suggestions
  • One-click replay to validate safer prompts

📊 Simple AI Observability Dashboard

  • Live metrics for:
    • Latency
    • Token usage (cost awareness)
    • Error rate
    • Risk level
  • Beginner-friendly visualizations
  • Automatic alerts when behavior looks abnormal

🎨 Clean, Student-Friendly UI

  • Chat-style interface inspired by tools students already use
  • Real-time feedback and notifications
  • Designed to feel approachable, not intimidating

🛠️ How We Built It

Tech Stack

Frontend

  • Next.js
  • React
  • Tailwind CSS
  • Datadog Browser Logs SDK

Backend

  • Node.js
  • Express
  • Google Vertex AI (Gemini Flash)
  • Datadog APIs (logs, metrics, incidents)
  • StatsD for low-overhead metrics

🔄 How It Works

  1. A user submits a prompt in their AI application
  2. AegisAI analyzes the prompt for security and risk
  3. Safe prompts are sent to Gemini
  4. Metrics and logs stream in real time
  5. Risky behavior triggers alerts with explanations
  6. The AI response returns with helpful metadata

All monitoring runs live, not simulated.

🏆 What Makes AegisAI Different

Most hackathon projects rely on mock data or static demos.

AegisAI is real.

  • Real Gemini API calls
  • Real security detection logic
  • Real metrics and incidents
  • Real traffic simulation
  • Real-time dashboards

You can clone the repository, run one command, and immediately observe live AI behavior and safety signals.

📚 What We Learned

  • AI systems fail differently than traditional software
  • Tokens directly translate to real costs
  • Security risks are contextual, not binary
  • Beginners need explanations, not just alerts
  • AI observability should be educational, not intimidating

🚧 Challenges We Faced

  • Reducing false negatives in prompt detection
  • Designing alerts that are helpful without being overwhelming
  • Managing API rate limits during testing
  • Balancing technical depth with beginner usability

Each challenge resulted in a more robust and accessible system.

🎯 Accomplishments

  • Fully working end-to-end AI monitoring system
  • Real-time safety detection and explanations
  • Production-grade observability pipeline
  • Clean, beginner-friendly interface
  • Designed for learning and experimentation

🚀 What’s Next

Short Term

  • Beginner “learning mode” with guided explanations
  • Slack / Discord alerts for student teams
  • Multi-model support

Long Term

  • Open-source detection rules for students
  • AI red-teaming playground for learning security
  • Community-driven AI safety education

🌱 Built For CodeSpring Hackathon

AegisAI was refined and repositioned during CodeSpring Hackathon
to support students and creators building AI responsibly.

🔗 Links

TL;DR

AegisAI helps students build safer AI apps by detecting risks, explaining failures, and guiding fixes in real time.

This is not just a tool.
It is a learning safety net for the next generation of AI builders.

Share this project:

Updates

Submission history