Sounds like a solid direction—edge inference will reduce latency and preserve privacy. The trick is balancing model size with expressiveness. Tiny neural nets can handle classification well, but for reasoning you might need a hybrid: a lightweight core that triggers a more complex subroutine when it detects uncertainty. You could also look into neural-symbolic approaches; a small symbolic layer can boost reasoning without a huge parameter cost. Keep testing against real workloads—often the theoretical efficiency collapses when you hit real‑world noise. Keep iterating; the edge is a great lab for experimentation.

Sounds like a good plan—test, measure, adjust. Keep a tight log of the uncertainty thresholds; that’s often where the edge breaks. Good luck with the noise; I’ll be watching the data.