Yug D Oswal

I am a third year UG CS student at VIT Vellore. I'm most recently working on a paper we're soon going to be submitting to a 2025 A* conference. I broadly do work in representation learning, theoretical and empirical deep learning. I'm most interested in exploring the intersection of mechanistic interpretability, representation learning, and the theory of deep learning. I specialize in AI/ML and am also proficient in data science, web dev, and app dev (Flutter).

I'm planning on pursuing an MS (extendable to a PhD). While I do have my interests, I'm still very interested in increasing my breadth by exploring other SOTA fields of research in AI/ML.

Personally, I enjoy dancing, music, and reading novels. I was an abacus champion during my school years. I like thinking up crazy theories for neural nets sometimes. I was also starred in Intel's Developer Spotlight for my project Rekindle.
Currently (still) working towards the best 'me' I could think of.

Email  /  Resume  /  Google Scholar  /  LinkedIn  /  GitHub

Machine Learning Research

Computing hyperstrips instead of hyperplanes in euclidean space. 4.6% accuracy gain on ImageNet, with 46.4% parameter reduction in VGG19. Enables smaller yet performant foundational models.

Proposed SMAE and SQRT, MAE-like differentiable losses, which are robust to noisy input data and natural distribution shifts. Proposed novel classification losses and a customized loss scheduling algorithm for the same.

Experience

Developed the primary prototype for an anti-UAV system involving object detection and tracking, along with a web app and a unique migratable deployment in isolated systems.

Referred by Raghu Bala sir, a MIT AI Course Facilitator and Founder. Worked on developing and deploying pipelines such as agentic workflows, RAG, agentic tooling, chat history context aware, and guardrails for fine-tuned LLMs used in humanoid speech-capable autonomous agents serving de-addiction therapy. Also worked on classical ML POCs for clients in the insurance domain.

Led an international project team and coordinated with mentor, team, university, and client. Worked on Named Entity Recognition, Geocoding, Incremental Clustering, and engineering an optimized comprehensive pipeline. Resolved 5 real-world issues for Signal Corp Ltd.

Projects

Developed a service to aid Alzheimer's and Dementia disease patients. Trained an emotion extraction model incorporating my novel loss function research and deployed local LLMs. Patients can create life-journals and relive memories based on vague remembrances of emotions and snippets of events they have spoken to the Rekindle companion. Currently being refurbished.

Independently developed the backend for the Computer Society of India event website, Lasertag, held during the college fest Gravitas. Used by over 1000 students and developed using Node.js, Redis, MongoDB (Atlas), and integrated with CI/CD.

Community or Volunteering

I am the Research and Development Head of the Computer Society of India chapter at VIT. My work involves mentoring juniors, organizing and managing events, providing research opportunities, guiding research, and directing chapter activities and future directions.

This website was produced from a template made by Jon Barron.