-
-
HOW IT WORKS
-
HOW IT WORKS
Inspiration
Most African users face real cybersecurity threats daily phishing links shared on WhatsApp, fake banking sites, scam URLs with no reliable tools to check them. Existing solutions require internet, cloud APIs, or expensive hardware. AURA was built to change that.
What it does
AURA is a fully on-device autonomous cybersecurity agent. You give it a URL, it reasons about it, calls its HAsH Guardian ML skill, and returns a threat verdict all without touching the internet.
How we built it
- Local LLM (qwen2:1.5b) via Ollama as the reasoning brain
- Custom-trained RandomForest phishing detection model (ROC-AUC 0.9113, trained on 549K+ URLs, 65 features) as an in-process skill
- Python + FastAPI architecture with skill-based tool system
- Runs on a ThinkPad T430 (Core i5-3210M, 8GB RAM, Intel HD Graphics 4000)
Challenges
- Stripping all cloud dependencies from an agent originally built on Gemini/Groq
- Making the ML model run standalone without the full HAsH Guardian web stack
- Getting everything working on CPU-only hardware with no GPU acceleration
What we learned
A small local LLM doesn't need to be smart, it needs to be a good orchestrator. The tools do the heavy lifting. This architecture works on any commodity hardware.
What's next
Expand AURA's skills to cover malware file analysis, network traffic inspection, and real-time threat feeds all offline.
Log in or sign up for Devpost to join the conversation.