Dev Gupta

I am a software engineer currently working at FTAI Aviation as an AI/ML specialist. I used to be an undergraduate researcher working at the Multimodal Vision Research Lab at Washington University in St. Louis.

I received my B.S. in Computer Science & Mathematics at WashU.

Email  /  LinkedIn  /  GitHub  /  HuggingFace

During my research, I gained expertise in developing state-of-the-art machine learning models and optimizing AI inferences for scalability and performance. Successfully engineered a multimodal model and retrieval system utilizing dense vision embeddings and developed a cross-vision pose estimation model that conditions on 3D reconstructions. Eager to apply technical skills and innovative solutions in advanced AI projects. Passionate about ML/DL/CV ∪ NLP ∪ Embodied AI

GeoDiT is a diffusion transformer for text-to-satellite image generation that replaces pixel-level conditioning with a simpler point-based control scheme and adaptive attention, enabling more flexible and semantically rich control while outperforming prior remote sensing generative models.

The work extends cross-view pose estimation from single images to multi-view inputs by adapting a multi-view 3D reconstruction framework and training on aligned ground-to-aerial datasets to jointly estimate 3DoF poses more robustly in aerial reference frames.

This project proposes a training-free stereo video generation pipeline that combines depth estimation, latent-space warping, and diffusion-based video synthesis to generate geometrically and temporally consistent right-view videos from monocular left-view inputs while exploring cycle-consistency constraints to improve fidelity.

Quoted by WashU's The Source - Read Article

• DI2 Project Presentation
• DI2 Project Report