Mantha Sai Gopal

Machine Learning Research Engineer

Working on Computer Vision and Representation Learning.

I am an ML Research Engineer at CamCom Technologies, where I develop computer vision systems that bridge advances in machine learning research with real-world deployment. My work focuses on 3D vision, scene understanding, vision foundation models, representation learning, and training-free adaptation methods that enable large pretrained models to solve new tasks through in-context learning rather than additional optimization. I hold master's degrees in Computer Science and Mathematics, and my mathematical background continues to shape how I approach machine learning research, emphasizing first principles, geometric reasoning, and rigorous analysis. I am particularly interested in understanding how rich visual representations can be leveraged to build robust and scalable perception systems for challenging industrial applications. This website collects my research, technical writing, curated reading notes, and explorations across computer vision, representation learning, and mathematics. Beyond my research, I remain a hobbyist mathematician with a longstanding interest in number theory, drawn to the elegance of its ideas and the joy of exploring problems for their own sake.

News

Jul 2026

Published our preprint on arXiv: REBASE: Reference-Background Subspace Elimination for Training-Free In-Context Segmentation.

May 2026

Paper accepted at the 4th International Conference on Recent Advances in Applied Mathematics (RAAM 2026): A Probabilistic Framework for the Erdős–Kac Theorem.

Research & Writing

Writing
DINOv2 Paper Explained
A deep dive into DINOv2 and self-supervised representation learning for vision transformers.
Writing
DINO Paper Explained
A detailed walkthrough of DINO, self-distillation, and self-supervised learning for Vision Transformers.
Research
REBASE
Training-Free In-Context Segmentation Using Background-Subspace Elimination.
Research
A Probabilistic Framework for the Erdős–Kac Theorem
Accepted at RAAM 2026.
Writing
YOLOv1 Paper Explained
A detailed walkthrough of YOLOv1, the first unified real-time object detection framework.