CV

Ph.D. in Industrial Engineering

• Research in efficient adaptation and knowledge compression of large pretrained vision, language, and multimodal models.
• Efficient world models and vision-language-action (VLA) models for embodied agents (since Aug 2026).
• Continuing DistillLens (symmetric knowledge distillation for LLMs) and DistillHPC (label-free distillation for HPC code generation, with Oak Ridge National Laboratory).

M.S. in Computer Science

• Admitted to the Ph.D. program in Computer Science; earned the M.S. en route before transferring to UTK with the advisor's lab in 2026.
• Deep learning projects on 2D/3D computer vision and large language models (LLMs), funded by the US Department of Defense (DoD).
• Started DistillLens (Oct 2025), a symmetric logit-lens distillation method for compressing LLMs, and DistillHPC (May 2026), label-free distillation for scientific code generation in collaboration with Oak Ridge National Laboratory (ORNL).

Bachelor in Computer Engineering

• Ranked 11th in the engineering entrance exam, competing with 15,000+ candidates, received full scholarship for undergraduate study.
• Gained knowledge about significant CS courses like AI, Image Processing, Data Structure & Algorithm, DBMS, Software Engineering, and so on.
• Thesis: Automatic speech recognition for low-resourced Nepali language which was later presented at an IEEE conference.

DistillLens

• Symmetric knowledge distillation that projects teacher and student intermediate hidden states into vocabulary space via the logit lens and aligns them with a symmetric divergence.
• Outperforms standard distillation and feature-transfer baselines across GPT-2, Llama-2, and Qwen model families on instruction-following and math reasoning benchmarks, with no added inference cost.

DistillHPC

• Label-free knowledge distillation to train compact coding assistants for the parallel programming languages used on DOE supercomputers, without human-authored expert examples.

Efficient World Models and VLAs

• Applying efficient adaptation and knowledge compression to world models and vision-language-action (VLA) models, so one shared model can be specialized to many tasks from limited demonstration data.

Graduate Research Assistant

• Research on efficient world models and vision-language-action (VLA) models for embodied agents.
• Continuing DistillLens and DistillHPC, the latter in collaboration with Oak Ridge National Laboratory (ORNL).

Graduate Research Assistant

• Deep learning projects on 2D/3D computer vision and large language models (LLMs), funded by the US Department of Defense (DoD).
• Developed parameter-efficient fine-tuning for point clouds (GFT, WACV 2026) and knowledge distillation methods for LLMs (DistillLens; safety-alignment distillation, EMNLP Findings 2026).

Research Assistant

• Developed skills for object detection and segmentation tasks on 2D medical images and explored their multi-modal approach ( esp. vision-language models ); also worked for segmentation with 3D mesh data.
• Demonstrated strong skills in writing scientific manuscripts, with multiple papers submitted for review, showcasing the ability to communicate methodologies, results, and implications effectively.
• Ensured reproducibility and modularity in ML projects by implementing robust methodologies and practices, allowing for the transparent and replicable programming of the projects.

Trainer

• Training a group of interns to develop medical imaging applications with the use of ML.
• Instructing and guiding them about ML through activities like paper reading sessions, lab and lecture sessions, and topic presentations.

Teaching Assistant

• Guided participants through a series of labs related to neural networks, transformers, federated learning, graph neural networks, active learning, and so on.
• Mentored three groups during Hack-a-Dev, a 10-day machine learning hackathon.

Programming Instructor

• Provided tutoring on software development life cycle and assisted participants with software documentation and library/framework installation.
• Taught participants about API development for web applications, emphasizing its concepts, best practices, and usage.

Lower Limb Segmentation

• Conducted training experiments of different deep learning models on the remote server to segment three bones – knee, pelvis, and ankle – from CT scans of the lower limbs of patients.
• Deployed a robust Python rest API on the remote server for the segmentation request from a client, with a pipeline including pre-processing, inference, and post-processing steps.
• Ensured interoperability, reproducibility, and understandability of the deployed application using Docker, and well-structured documentation and comments.

Vision Language Segmentation Models (VLSMs) for Medical Images

• Reported zero-shot and finetuned segmentation performance of 4 VLSMs on 11 medical datasets using 9 types of prompts derived from 14 attributes, prompts are given as text conditioning informations.
• Worked with encoder-decoder architecture to generate binary segmentation masks for VLSMs.
• Tested the compatibility of the VLSMs (such as CLIPSeg and CRIS) pre-trained for open-domain images with medical images.

Object detection in 2D Orthopantomogram (OPG) Images

• Critically analyzed the literature and state-of-the-art models for different segmentation and detection tasks on radiology images of dentistry and their inadequacy.
• Designed and developed the data annotation tool for object detection over 2D OPG images.
• Working on identification and localization of dental anatomical structures and abnormalities while benchmarking with existing methods like YOLO, RetinaNet, RCNN, and FastRCNN.

Segmentation in 3D Teeth Scan

• Learned about the representation and preprocessing of 3D mesh and point cloud data.
• Benchmarked with different 3D point cloud segmentation models such as Pointnet/++ and DeltaConv.

SR-GAN

• Open source implmentation of SuperResolution using GAN (SRGAN) in Keras framework.

Nepali AutoComplete and LM

• Open source implementation of Nepali AutoComplete and Language Model using LSTM in PyTorch framework.

SahaYatri

• Hackathon project to promote toursim in Nepal developed using Django and ReactJS.

• AAAI 2027 Reviewer (invited; 4 papers reviewed).
• MICCAI 2026 Reviewer (invited; 4 papers reviewed).

• AAAI 2026 Reviewer (invited; 4 papers reviewed).

• LMIC Travel Grant by MICCAI.
• Presidential Fellowship within TCV initiative by GSU (Only 6% of the doctoral students).

• Undergraduate funded by Government of Nepal.