I build deep learning models that make sense of medical images — teaching computers to see tiny abscesses, nerves, and tissue boundaries that are easy to miss.
Built a SAM-Med3D refinement pipeline using nnU-Net predictions as point prompts. nnU-Net and SAM-Med3D land on similar accuracy overall (~0.60–0.63 Dice) — but they fail in totally different ways, which turns out to be the interesting part of the story.
Added a trainable CRF layer to a ViT segmentation model to clean up predictions (+10% Dice on BraTS2020), plus an unsupervised twist using graph cuts on attention maps.
Trained a 3D CycleGAN to translate T1→T2 MRI scans, then segmented tumors across mismatched scan types without any paired labels (0.40 Dice).
Built a custom decoding trick that hides an invisible multibit watermark in generated text — and actually made the text ~10% less perplexing, not more.
A hackathon app that looks at a photo and composes matching music, mixing LLaVA and MusicGen behind a simple Gradio UI. Built at TikTok TechJam.
A Canva App that turns a text prompt into a clean-cut sticker using diffusion + background removal. Built at Canva's AI Integration Hackathon.
A 3D Unity game controlled by real muscle signals, built to help cerebral palsy patients make rehab exercises feel like play. Built at CWRU Hackathon and won 2nd place.
