Where I've built —
from Cornell-backed
startups to competition robots.
Three roles across robotics, education, and B2B SaaS.
Each one taught me to move from prototype to
deployment on real hardware and real users.
Nova Robotics
Things I've built — shipped
to real hardware, real teams.
A selection of side work spanning open-source ML, edge AI, and competitive robotics. Full codebases available on request.
VibeVoice × Rokx
Contributed to Microsoft's open-source VibeVoice (neural TTS) — enhanced emotional expressiveness and added real-time inference. Fine-tuned a quantized LLaVA VLM with input from a Stanford childcare professor and deployed the full stack on a Jetson Nano at 15+ FPS locally. Dockerized the omnimodel stack — hours-long setup down to minutes, LAN or EC2 deploys.
EDGE
$ docker run rokx/vibevoice:latest
+ loading LLaVA q4_k_m · jetson-nano
+ inference live @ 15.7 fps · 97ms/token
EdgeAI Tennis Motion Capture
Low-cost racket motion capture for 20+ players. Kinematic data logged to SD for batch training; streamed over Wi-Fi for live feedback. Custom CNN + LSTM hit ~90% stroke classification accuracy with inference offloaded to mobile.
FTC Robotics · Team #8872
Programmed a robot ranked top 1% globally — placed 36th of 7,681 teams worldwide. Directed a 5-member software team through season cycles.
The Stack
Formal coursework at Mission College (all A's), self-taught everywhere else.
Building something hard?
Let's make it real.
Open to firmware, robotics, and applied-ML roles —
internships for Summer 2026 and beyond. I reply to every
message, usually within a day.