My projects usually begin with a failure mode in a distributed or resource-constrained learning system, then connect theory, attacks, algorithms, and implementation.
Continuous audio representation learning for heterogeneous edge devices.
StreamSplit replaces large negative-sample memories with a compact distributional representation and dynamically chooses where to split computation between edge and cloud using resource and uncertainty signals.
Information-theoretic and empirical work on what intermediate activations and gradients can reveal, when standard privacy calibration fails, and how client heterogeneity changes the guarantees.
Examining attacks that can appear benign in gradient space while remaining anomalous at the split representation, motivating spectral auditing directly in activation space.
Real-time causal forensics for cross-slice attacks, explicitly modeling shared-resource confounding and certifying statistical validity, robustness, and privacy properties.
Separating sensitive information already present in model inputs from demographic signal added or amplified deeper inside a language model.
Federated and privacy-preserving learning for smart infrastructure, including waste classification under decentralized data ownership.
I’m increasingly interested in hardware-aware ML, efficient audio intelligence, quantization, and lightweight multimodal interfaces for constrained devices.