Class-agnostic obstacle detection for autonomous surface vessels on the unified USVTrack + LaRS benchmark, with synthetic IMU horizon stabilization and Kalma...
A CS5330 teaching video building up the Transformer from tokenization to self-attention to the MLP block, then porting the whole architecture to images with ...
A CS5330 presentation on my favorite figure in computer vision
An overview of the Fourier Transform
A 30-layer CNN I built from scratch vs. ViT-B/16 on brain MRI for 4 types of brain tumors.
A miniature self-driving perception stack on 70K BDD100K dashcam images. RT-DETR for detection, U-Net (from scratch) for drivable area, MiDaS for depth. All ...
A solo-built, production-deployed platform hosting my AI/ML work under one roof — brain tumor classification with CNN vs ViT interpretability, an autonomous ...
Conditional GAN for face generation across 7 ethnic categories. Per-ethnicity FID variance <3% and human perceptual variance <8% — no significant ethni...
A 100-day self-designed ML curriculum in three phases — data foundations, classical ML, deep learning — with capstones between phases. Currently at 67/100
Three CNN architectures benchmarked across pneumonia, tuberculosis, and lung cancer detection from chest X-rays and CT scans. Sequential won the X-ray tasks ...
