Architected a scalable deep learning framework for automated passive sonar classification, directly enabling NPOL's underwater acoustics surveillance capabilities.
Engineered a memory-efficient data pipeline to systematically chunk high-resolution audio files into compressed .npz tensors, mitigating cloud Out-of-Memory (OOM) bottlenecks during scale-up.
Scaled the DWSTr deep learning framework to process the massive 47-hour DeepShip real-world acoustic dataset, targeting the classification of 265 distinct vessels across 4 commercial categories.
Presented the model to senior leadership, facilitating high engagement Q&A to gather actionable feedback for iterative model enhancements
Built the initial end-to-end audio pipeline processing 10,000+ sonar samples into Mel-spectrograms, reducing preprocessing latency by 35%.
Developed the foundational hybrid DWSTr (Depthwise Separable CNN + Transformer) model, achieving a benchmark 97.17% accuracy across 12 vessel categories by implementing square root class-weight smoothing to resolve severe dataset imbalances.
“Your love makes me strong, and your hate makes me unstoppable”