Developers Build Dual-Model AI System to Detect Urban Sound Events in Real Time

A team built SonicSentinel AI, a sound-event detection system that runs two independent machine learning models — one trained in Python and one trained through a separate platform — to classify urban audio events such as gunshots, screams, and alarms. The dual-model design creates a natural ensemble, where agreement between models boosts confidence and disagreement triggers caution. The dataset was built from scratch with roughly 300 audio clips per class, and the best-performing Python model, an XGBoost classifier, achieved a test accuracy and macro-F1 score of 0.693, falling short of the team's own target floors of 0.85 and 0.80 respectively. The system includes audio quality gating, an alert state machine, and configurable JSON-based rules so operators can adjust behavior without modifying code. Inference runs at 0.81 milliseconds per clip on CPU, well within the system's 8-second response allowance.