Explore a selection of my published work and research across different areas of professional interest. While a lot of what I've worked on isn't publicly available, I hope you enjoy what I'm able to share here.
NOTE: Everything here represents my personal thoughts and opinions, not those of any company I've worked with.
Leveraging AI to search petabytes of Earth observation data is a massive challenge, but it’s exactly what our team at BlackSky is focused on. This post includes only information that has been publicly released by the company, given the sensitive nature of the work.
Smartflow is a cloud-based framework we developed at BlackSky to enable scalable spatiotemporal and geospatial AI research. Built on open-source tools and Kubernetes, it streamlines the processing of large volumes of Earth observation data and makes model experimentation more efficient and scalable.
Training high-performance object detectors that work at a global scale is no small feat—but it’s exactly what our team at BlackSky is focused on. This post includes only information that has been publicly released by the company, given the sensitive nature of the work.
We developed machine learning models to estimate road transportation emissions using satellite imagery, starting with the United States where reliable inventory data is publicly available. Once trained, these models can be applied globally, with appropriate consideration for regional differences.
We introduce a deep learning approach for estimating and localizing population counts in displacement camps using high-resolution overhead imagery. Trained on census data from refugee camps in Cox’s Bazar, Bangladesh, the model achieved strong accuracy on unseen data. This work is a step toward building tools that can help the humanitarian community respond more effectively and rapidly to global displacement challenges.
We explored the use of machine learning to identify exoplanets using data from the Gemini Planet Imager, combining real and synthetic observations to train our models. While we were able to build models that detected some promising signals, they ultimately did not generalize well to withheld data. Nonetheless, the project provided useful insights into the challenges of applying AI to direct imaging in astronomy.
We introduce the Functional Map of the World (fMoW) dataset, which includes over a million satellite images from more than 200 countries, labeled with functional building and land use categories. The dataset is designed to support machine learning research that combines visual data with rich metadata such as location, time, and sun angle. With bounding box annotations, temporal views, and baseline models, fMoW aims to serve as a foundational resource for geospatial AI in the same way ImageNet did for computer vision.
We developed an automated approach for segmenting brain structures such as CSF, white matter, and gray matter in MRI scans. Our method combines classic image processing techniques with machine learning to improve accuracy.
We developed a patented machine learning approach for automatically detecting pneumothorax in ultrasound. The method first identifies the pleural interface, then analyzes pleural sliding patterns to distinguish normal from abnormal motion that may indicate a pneumothorax.
We explored several methods for segmenting the heart’s inner walls from 3D transesophageal echocardiography (TEE) to support diagnostics and biomechanical modeling. Our approaches include a level set method, a graph cut technique using radial symmetry, and a random walker method—all designed to automatically delineate the endocardial surface.
We built a preoperative planning tool to help model mitral valve closure using patient-specific anatomy from 3D transesophageal echocardiography (TEE). The approach uses physics-based modeling and a user-guided segmentation step to predict how well the valve leaflets will come together.
We extended a popular 2D optical flow method (Brox et al.) to work with 3D ultrasound data—specifically transesophageal echocardiography. By combining classical ideas with modern techniques, we were able to track heart motion more accurately than current methods.