Molecular generation

Built the original React molecular-design platform from the ground up, then developed and deployed a conditional molecular generation model. The platform was subsequently used for 1,000+ virtual screens, with 700+ generated compounds tested across eight programs.

Structure prediction

Worked on diffusion-based molecular docking and water-aware Pearl, owning data-pipeline and model updates, loss changes, and optimization. Coauthor of the Pearl arXiv preprint.

Predictive models

Developed GNNs for molecular-property prediction. Improved multitask ADME models, reducing aggregate prediction error by 16% across nine endpoints on an external benchmark.

Research tools

Built a React, Flask, and PostgreSQL experiment dashboard used daily by researchers. Made large experiment histories load an order of magnitude faster.

Developed a deep learning model for monocular 3D object detection and tracking using dashcam videos and phone-sensor data to analyze drivers’ risk behaviors. This research formed my M.Eng. thesis through MIT’s 6-A program.

Supervised by Prof. Frédo Durand. Researched automated shape collection of car models and 3D shape segmentation on a custom dataset. Built tools for converting between 3D representations and visualizing segmentation labels.

Worked in the Tools and Compilers and Trading System groups. Improved the internal version-control and code-review tool by eliminating high-overhead calls, and built a testbed to measure trading-algorithm latency by simulating trade orders. Implemented both projects in OCaml.

Improved artistic style-transfer speed by 2.1× through neural-network architecture changes. Researched transfer learning for image classification and object detection in TensorFlow.

M.Eng., May 2020 · Concentration in Artificial Intelligence.

B.S., June 2019 · Electrical Engineering and Computer Science.

Teaching assistant for Mobile and Sensor Computing in Spring 2019 and Spring 2020.