Research

Self-supervised representation learning for sensor networks, time-series forecasting, and robust systems.

Research Statement

My research focuses on self-supervised representation learning for real-world time-series and sensing data, particularly in domains where precursor detection is critical: environmental flood warning, glacial lake stability, and anomaly detection in distributed financial networks.

I approach modeling through a systems-oriented lens: designing ML pipelines that remain computationally efficient under dynamic workloads without assuming massive cluster availability.

A central pillar of my methodology is research reproducibility and falsifiability. I structure experimental workflows around pre-registered protocols and report both positive performance and negative failure bounds.

Core Areas & Interests

Self-Supervised Representation Learning Time-Series & Sensor Dynamics Environmental Early-Warning Systems Digital Signal Processing Topological Data Analysis (TDA) Real-Time & Elastic ML Infrastructure Reproducible ML Frameworks

Active Research Threads

1. flood-sentinel: Temporal Encoders for Hydrological Sensing

Self-supervised representation learning on high-frequency USGS streamgage observations to isolate precursor signatures of flash floods.

[Project Details]

2. BPFeat: Backpressure-Driven Elastic Feature Windows

Adaptive feature extraction mechanisms designed to regulate latency under variable data streaming rates in edge and server setups.

[Project Details]

3. sentinel-gl: Remote Sensing for Glacial Lake Hazards

Masked autoencoding architectures applied to multispectral satellite time series for retrospective GLOF risk evaluation.

[Project Details]

4. Topological Financial Network Anomaly Detection

Applying persistent homology computations over dynamic transaction graphs to uncover coordinated collusion structures.

[Project Details]