Accomplishments that we're proud of

🛡️ Guardian Agent Breaker (GAB)

💡 Inspiration

As an Electrical Engineer specializing in hardware validation, I know that when an AI agent interacts with hardware, a hallucination isn't just a wrong answer—it’s a physical failure. In the rush to deploy agentic workflows, we are giving LLMs "hands" (APIs) without the physical-layer circuit breakers we’ve used in engineering for decades. I built GAB to ensure AI autonomy never results in a blown capacitor or a fried FPGA.

🏗️ How I Built It

GAB is a Deterministic Interception Layer that sits between the AI and the hardware controller.

  • Hardened Supervisor: Built using Node.js inside a Chainguard container. This ensures a zero-vulnerability execution environment, preventing the supervisor itself from being compromised.
  • Risk Brain: Powered by PostgreSQL (via InsForge). Safety parameters are stored out-of-band so the agent cannot "convince" the system to rewrite its own safety rules.
  • Auditability (Guild.ai): I integrated Guild.ai to treat every intervention as a tracked experiment. This provides an immutable forensic record of why an agent was blocked, approved, or warned.
  • Real-time Interception: Uses Redis for state management and rate-limiting to prevent agentic "loops" from flooding hardware controllers.

🧠 What I Learned

Safety must be binary, even when AI is probabilistic. Integrating a hardware-abstraction layer into an AI workflow proved that we can maintain the flexibility of LLMs while enforcing the rigid safety standards of traditional engineering. I also mastered container hardening with Chainguard to meet industrial security requirements.

⚠️ Challenges I Faced

The primary challenge was the Latency vs. Safety trade-off. To act as a true "circuit breaker," GAB must intercept intents in milliseconds. Optimizing DB lookups and Docker network bridging to Redis was critical. Furthermore, I had to develop a granular risk-scoring logic to differentiate between harmless telemetry ($V_{op} = 3.3V$) and catastrophic overvoltage commands ($V_{max} > 10.0V$).

🚀 Technical Impact

By isolating safety logic in a hardened container, GAB provides a Zero-Trust architecture for AI-to-Hardware interaction. It is a digital fuse for the next generation of autonomous industrial systems.

Built With

Share this project:

Updates