Dr. Kevin Decker

AI/ML researcher working on foundation models, self-supervised pretraining, and representation learning for sensor and trajectory data

I am an AI/ML researcher focused on foundation models and self-supervised pretraining for data that falls outside the natural image domain. My work spans SAR, LiDAR, hyperspectral, EO, and single photon imagery alongside large scale trajectory and time series data, with an emphasis on learning transferable representations in domains where no pretrained backbones exist.

I currently lead model architecture and pretraining strategy for a foundation model in continuous time trajectory forecasting, covering corpus construction, tokenization and input design, linear time sequence backbones, probabilistic forecast heads, and architecture search on multi node GPU clusters.

Across defense and research programs I have built scalable ML infrastructure, training and evaluation frameworks, and data curation tooling supporting rapid experimentation from research through operational deployment.

My doctoral research at the Air Force Institute of Technology focused on multimodal remote sensing fusion and statistically sound evaluation methodologies for spatial data.

Key Skills

Foundation Models and Pretraining

Self-supervised pretraining including masked image modeling and joint-embedding predictive objectives, transfer of pretrained encoders to downstream tasks, and pretraining corpus construction at terabyte scale.

Multimodal and Sequential Data

Technical expertise spanning SAR, LiDAR, hyperspectral, EO, radar, and single photon sensing, extending to irregularly sampled multi-source trajectory and time series data across maritime, air, and ground domains.

Architecture and Training at Scale

Transformer, state-space, and linear-attention architectures, distributed training and neural architecture search on multi-node GPU clusters, and probabilistic modeling for forecasting under uncertainty.

ML Infrastructure and Tooling

Modality-agnostic training frameworks, reusable evaluation harnesses, and data curation pipelines supporting rapid experimentation, model validation, and deployment workflows.