We are building a way to help make safer AI.We're building SignalGuard: a safety gate for agentic tool calls, running fully offline. A local LLM (Qwen3 via Qualcomm GenieX on a Snapdragon X Elite NPU) controls a physical Arduino UNO Q over MCP, but every tool call first passes through an independent, smaller classifier model whic is separate from the acting model so a jailbroken model can't also control the check meant to catch it.Anything malicious is hard-denied.
.png&w=256&q=75)