I built MicroSafe-RL to solve the "hardware destruction" problem during Reinforcement Learning and Edge LLM deployment.

The Tech: It’s a bare-metal C++ interceptor using an EMA+MAD stability metric derived from Control Lyapunov Functions. Performance: 1.18 microseconds worst-case execution time (WCET). No heap, no dynamic allocation, just 24 bytes of state. The "Bridge": The latest update includes a Python-C++ bridge to use local LLMs (like Gemma 4 via Ollama) as robotic controllers while keeping them physically safe.

Currently under review at IEEE Transactions on Aerospace and Electronic Systems.

GitHub: https://github.com/Kretski/MicroSafe-RL