Building production models for high-precision 3D reconstruction, pose prediction, and tracking of dental arches from images captured on mobile devices.- Delivered SmileMaker platform, an iOS app that reconstructed a potential customer’s dental arch in 3D from images captured on a phone, and displayed a 3D treatment plan for their participation with SmileDirectClub. SmileMaker was ranked the #3 Medical product on the iOS Store.- Helped develop and validate SmileMaker’s core 3D reconstruction algorithm, which paired a PyTorch3D renderer with a bundle adjustment process. It used a PyTorch Mask R-CNN-based PointRend model with ResNet backbone trained on customer photos to segment teeth for 2D-3D correspondence-building.- Adapted SmileMaker’s backend to process videos of camera sweeps across a user’s arch instead of photos, and generalized its frame selection pipeline to select an information-rich collection of video frames for use in 3D reconstruction. - Engineered a PyTorch regression that optimized through se(3) to predict 6DoF tooth pose based on a set of images; employed a PyTorch3D differentiable renderer for comparing images and 3D teeth, and regularized the model with a Gaussian Mixture Model. The technology was used in SmileMonitoring, a product which predicted whether customer’s tooth positions were off-track or on-track for their treatment.- Made regular contributions to the SDK, including an OpenCV-based optical flow tracking system that ran on videos, matrix and algebraic operations with SE(3), and a PyTorch3D-based SoftRas differentiable renderer.

Developing machine learning-based methods to quickly generate 3D content in Imbellus’ digital hiring assessments.- Built a product for 3D artists to automatically reconstruct animal meshes in 3D from a single image, reducing Imbellus’ mesh creation time by 2/3. - Performed 3D reconstruction using AtlasNet generalized to accommodate mesh priors and semantic segments, paired with a differentiable renderer for correspondence-building. Semantic segments were obtained by fine-tuning DeepLab on custom datasets for fine-grained image segmentation.- Textured 3D meshes using a Deep Image Prior network trained to hallucinate mesh textures in UV space based on a provided image.- Extended α-GAN’s architecture from 2D to 3D and trained it for 3D reconstruction on ShapeNet Core55. This work won the ICCV 2017 3D Reconstruction Challenge (https://shapenet.cs.stanford.edu/iccv17) and was presented at the ICCV 2017’s “Workshop on Learning to See from 3D Data.”- Built a 3D Free-Form Deformation model in Tensorflow for shape optimization; developed a regularizer on the Jacobian of the FFD to discourage self-intersections during optimization.- Promoted to team manager in 2019, and oversaw 3 direct reports, holding regular 1:1s and giving quarterly performance reviews.

Business travel optimization application, which generated optimal travel itineraries for business trips.- Led a team of 4 full-time engineers, and a global team of contractors.- Raised seed funding round at $1.6 million.- Developed an algorithm to pick an optimal itinerary from within a large space of itineraries subject to user’s preferences/constraints.- Helped build a Python backend on top of a Tornado web server and MongoDB (with Motor).

As part of the PhD program in theoretical physics, I worked with Alexander Voronov on categorification of topological quantum field theories. Withdrew before completion to found Jivry.

Studied Mellin representations of high energy scattering in AdS-space in order to probe conformal symmetry with Chung-I Tan.

Worked for Greg Landsberg (2009) and Menaakshi Narain (2010) with the CMS collaboration at the Large Hadron Collider. Performed statistical analyses on large data sets in python.