Bouncer - Google DeepMind x Cactus Compute Global Hackathon
AI Tinkerers - San Francisco
Hackathon Showcase

Bouncer

Team consisting of an Adobe ML Engineer (USC MSCS) and a Microsoft AI Security TPM, skilled in PyTorch, agentic LLMs, fine-tuning, and secure GenAI.

2 members Watch Demo
https://drive.google.com/drive/folders/1WBFKa-LdIlXO0g01rG_sjkYnWvh-ejLw?usp=drive_link

Aegis is an intelligent local privacy and security layer that enforces “local-first” sensitive data inspection for AI agents. Instead of sending user files straight to the cloud, Aegis acts as a semantic firewall. It intercepts agentic file requests and processes them entirely on-device using Cactus Compute to ensure speed and prevent sensitive data from leaving your machine.

Our architecture leverages a 2 model local pipeline: SmolLM2-1.7B first summarizes raw content, which is then analyzed by a SFT FunctionGemma-270M to classify sensitivity. Based on this on-device classification, our intelligent routing logic dynamically decides the next step:

  1. Pass safe content to Gemini 2.5 Flash for high intelligence reasoning.
  2. Sanitize PII locally before cloud transfer.
  3. Block critical secrets (API keys) instantly.
  4. Escalate ambiguous cases to the user for review.

This approach allows us to utilize Gemini without compromising users’ privacy.

1. FastAPI & Next.js: For UI/UX 2. TypeScript Middleware: To intercept agent tool calls at the system level. 3. SmolLM2-1.7B Cactus Compute Google DeepMind