Architected DesignPlanner, a 15k+ LOC PySide6 application to configure and launch the first phase of test-chip design. Built on a clean Model-View-Presenter architecture across UI, service, and repository layers with 95% test coverage. Built a configurable framework to run jobs locally or on a distributed cluster, with multithreaded live status updates to UI. Designing an agentic layer that reads STA results and auto-recommends constraints to improve chip performance.

Proposed and led creation of an internal MCP Server with 15 tools to automate lesson insertion into the AoPS platform. Secured the server with JWT auth for both internal (Claude Desktop, whitelisted IPs) and external (OpenAI API) access. Cut lesson input time from a 20 hr tedious manual process down to 2 hrs; deployed via Docker on Kubernetes.

Created longform data collection script for the ALOHA robot, increasing efficiency 4x (from 0.25 → 1 data pt/min). Streamlined filtering process for remotely stored (AWS S3 storage) behavior cloning data using a Streamlit webapp.

Created CNN in Python identifying 9 skin cancer lesions using TensorFlow, achieving 79% accuracy from scratch. Pivoted to lead ML research on distinguishing high/low risk melanomas (in hopes of replacing a competing $7,000 test). Adapted Harvard CLAM repo for WSI patching, segmentation, and feature extraction; 91% accuracy / 97% AUC.