computer engineering @ NC State
I trained a Unitree G1 in Isaac Lab to pick objects off the floor with PPO and no demonstrations. After the first policy gamed my success metric, I rebuilt the evaluator around contact physics; the new policy succeeded in 229 of 300 trials that began near the object. The full walk-to-grasp chain remains unsolved.
Python · PyTorch · Isaac Lab · RSL-RL
I attached an 83M-parameter side transformer to a frozen 4B code model and built the PyTorch training and evaluation system around it. The latest completed run reduced held-out next-token loss by 5.6% without changing the base weights, but has not improved end-to-end coding yet.
Python · PyTorch · Hugging Face Transformers · CUDA
I built and shipped a cross-platform desktop app around the Claude Code CLI, with persistent terminals, project switching, agent status, and Git diffs. Moving session indexing into Rust cut a 1,848-session benchmark from 2.8 seconds to 0.7.
TypeScript · Electron · React · Rust
I shipped an iOS news app and built its React Native and Supabase pipeline to ingest coverage from more than 30 publishers, cluster overlapping stories, and produce one multi-source brief per event.
TypeScript · React Native · Expo · Supabase
CacheForge adapts the method from Google DeepMind's FunSearch paper to discover CPU cache algorithms. An LLM generates C++ replacement policies, compiles them into ChampSim, and benchmarks them on SPEC CPU traces; measured IPC and a retrieval memory of prior experiments guide each new generation.
Python · C++ · ChampSim · SQLite · Groq · GPT-4o
I trained a CARLA driving policy from a semantic camera and vehicle state using PPO and behavior cloning, then connected it to ROS 2. Its best evaluation episode completed 33% of an urban route; full-route driving remains unsolved.
Python · PyTorch · Stable-Baselines3 · CARLA · ROS 2
I adapted LLM-guided program search to implied-volatility surfaces. Generated Python fitters are sandboxed, tested on 15 real and synthetic datasets, and rejected for calendar or butterfly arbitrage; across 96 candidates, the best beat ten classical baselines with zero violations.
Python · NumPy · SciPy · SQLite · Groq
I built a C++20 and Python options engine with a lock-free ring buffer, pybind11 analytics, and IBKR execution. The systems path benchmarked at 29 million messages per second and 16.7× over Python; the trading backtest remained too small to call conclusive.
C++20 · Python · pybind11 · CMake · Interactive Brokers API
I built a Mac-local iMessage agent that turns unanswered questions, promises, and IOUs into a SQLite ledger and sends a daily digest to your own thread. Its outbound adapter has no recipient argument, so the agent cannot message anyone else.
TypeScript · Bun · SQLite · Photon iMessage Kit · Groq
I built an interactive Starlink pass simulator where network behavior emerges from orbital geometry, phased-array loss, thermal and power state, RF, and retransmissions. Injected hardware faults propagate through the chain, and each dropped packet reports its physical cause.
TypeScript · React · Vite · satellite.js · Recharts
I built and deployed a web app that extracts assignments and exams from a syllabus PDF, lets the student correct them, and syncs the selected dates to Google Calendar.
TypeScript · Next.js · PostgreSQL · Groq · Google Calendar API
I built a Next.js search product that fetches and ranks learning material from six APIs in parallel. Reworking its YouTube requests cut quota use 83%, raising daily search capacity from about 25 to 90.
TypeScript · Next.js · Supabase · Upstash Redis
For Nutanix's intern hackathon, I built local middleware that masks sensitive information before a request reaches an external LLM and restores it after the response. Our team placed third.
TypeScript · Next.js · Python · Flask · llama.cpp
I built a dependency-free browser optical bench for Malus's law, with draggable polarizers, a live electric-field view, and two- and three-polarizer experiments rendered on Canvas.
JavaScript · HTML · CSS · Canvas API