Jonathan Tao

Morphi (HackIllinois 2026) — View on Devpost | View on GitHub

Python, Modal, vLLM, PyTorch, LoRA, Qwen3-VL, React Native

Morphi is an AI-powered sports coaching mobile app we built at HackIllinois 2026. Built with React Native, it lets users record short videos of their golf swings, basketball shots, or other physical activities, and get real-time, precision biomechanical feedback via voice narration and visual overlays.

This project was heavily focused on pushing the envelope of serverless AI infra and knowledge distillation. We needed the app to understand intricate human poses, but generic Vision-Language Models (VLMs) weren't precise enough for sports coaching. To bridge this gap, I designed a specialized data curation pipeline combining YOLO pose estimation with Gemini 3.0 Flash to synthesize over 2,300 expert-quality coaching examples. We then trained a custom LoRA adapter for a massive 32B parameter Qwen3-VL model on 8x H200 GPUs to teach the model how to correct specific joints based on proper sports mechanics.

To make the experience seamless on mobile, I engineered our backend on Modal. We utilized vLLM across a 4-GPU cluster to handle multi-modal inference efficiently, which slashed our full video processing and AI generation latency from over 3 minutes down to under 24 seconds, allowing for a tight, engaging user loop.

Custom CoreXY 3D Printer

Designed and constructed a high-performance CoreXY 3D printer from the ground up, featuring a rigid frame modeled in Autodesk Inventor and custom 3D-printed components.

Configured Klipper firmware on a Raspberry Pi 4 to drive salvaged Ender 3 electronics, achieving high-speed printing at over 1500 mm/s with 7000 mm/s^2 acceleration. The Video shown above is the printer printing a benchi.