µSentinel

µSentinel is a real-time intrusion detection system designed to run on the Renesas EK-RA8D1 evaluation board under µT-Kernel 3.0. The project was built for the TRON Project 2025 competition — I was one of a small number of sponsored competitors who received the hardware board after applying with a proposal.

The system combines two layers of detection. A rule-based baseline handles known signatures efficiently. On top of that, a lightweight autoencoder detects anomalies by learning the joint structure across system signals and flagging deviations through reconstruction error — catching the subtle, cross-signal intrusions that fixed thresholds miss.

The constraints were tight: inference under 5ms, memory footprint below 30KB. In testing against synthetic traces designed to reproduce realistic anomaly modes, the AI layer reduced false positives by around 40% relative to the heuristic baseline alone. I did not qualify for the second round of the competition, but the system was fully implemented and the results are reproducible from the repository.