Yug D Oswal

I'm a SPAR Fellow and final year CS undergrad at VIT Vellore, India. I'm most recently working on:

• Mechanistic interpretability (first year dream come true :) and AI safety with Shivam Raval, Harvard University, as part of the SPAR Fall 2025 fellowship.
• Creating the perfect dataset (note: not synthetic data but research in information theory and generative models) with Prof. Ravid Shwartz-Ziv from NYU.
  Fun Fact: He was Yann LeCun's postdoc (fanboy moment :)).
• Continuing ongoing work on statistical interventions for CoT alignment with Prof. Jindong Wang and Hao Chen.

My current research is around interpretability, safety, and catastrophic inheritance. I've previously worked (and continue) in theoretical/empirical ML and representation learning.

Please say hi! I love connecting with different people.

I'm planning on pursuing a MS (PhD subject to planning). Currently (still) working towards the best 'me' I could think of.

Email  /  Resume  /  Google Scholar  /  LinkedIn  /  GitHub

Machine Learning Research

Solved the computational bottleneck in higher order ANNs like quadratic neural nets (QNNs). Designed a reduced parameter (RP-QNN) and vectorized matrix forward/backward prop equations for RP-QNNs and QNNs.

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

Working on mechanistic interpretability and AI safety with Shivam Raval from Harvard as part of Fall 2025 fellowship. Developing control knobs to regulate CoT-output correspondence.

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.