Vijay Gohil

I’m an ML/CV Engineer at 2K Games and Visiting Researcher at the NYU AI4CE Lab. I work spans from building 3D computer vision systems with a focus on building systems that see and understand the 3D world to Vision Interpretability research.

My research interests include computer vision, sparse autoencoders for vision transformer interpretability, and multi-camera 3D reconstruction. I’m particularly interested in mechanistic interpretability for vision models — understanding what and how vision transformers represent the world internally.

Currently, I’m leading visaebench, a cross-architecture Sparse Autoencoder evaluation study training TopK SAEs on ImageNet across DINOv2-B, CLIP ViT-B/16, SigLIP, MAE ViT-B, and supervised DeiT, targeting NeurIPS 2026.

On the applied side, I work on object detection, pose estimation, and 3D triangulation, animation generation for Video Games— problems where geometric understanding meets real-time constraints.