Projects
Research Projects
SuTRA — Interspeech 2026
A structurally unified, morphology-aware tokenizer that preserves Indic akshara units and reduces morphological shattering.
Paper · Code · Project Page
HiDISC — NeurIPS 2025
A hyperbolic framework for domain generalization with generalized category discovery.
Paper · Code · Project Page · Slides · Poster
DG²CD-Net — CVPR 2025
An adaptive task-arithmetic approach to domain-generalized category discovery.
Paper · Code · Project Page · Presentation · Poster
Self Projects
AI Support Assistant (Samsung S25 Ultra Support)
Built a Streamlit-based chatbot integrating RAG with PDF manuals.
- Tech: Qwen3-4B-Instruct, QLoRA, RAG, Streamlit
- Fine-tuned Qwen3-4B-Instruct using QLoRA to adopt a patient and professional support persona.
- Delivered real-time, factual assistance with reduced hallucinations and improved user satisfaction by integrating RAG with PDF manuals.
Kinector: A Text-Conditioned 2D Gesture Generator
- Tech: PyTorch, MediaPipe, GRU
- Generated 2D stick-figure animations from text using a <100-sample short-video dataset.
- Built a pipeline with MediaPipe keypoint extraction, GRU pose prediction, and text-to-motion animation.
Course Projects (Computer Vision & GenAI)
Calorie Estimation from Food Images
- Built a multi-stage dietary-tracking pipeline using YOLO for food detection and GrabCut for foreground segmentation.
- Isolated food items from complex backgrounds to support calorie estimation.
Medical Image Deblurring
- Built a scale-recurrent network with spatial-asymmetric attention for multimodal medical images.
- Improved image clarity by 24% PSNR.
Learning with Noisy Labels using Vision Transformer (ViT)
- Classified CIFAR-100 images under 40% label noise using a Vision Transformer and the Turtle method.
- Achieved 83% accuracy while maintaining robustness to corrupted labels.
Fine-Grained Classification on CUB Dataset
- Designed a CNN with fewer than 10M parameters for fine-grained classification across 200 bird species.
- Achieved 87.14% top-1 accuracy with an efficient model design.
Achievements
- GATE 2022: Attained 97.84 percentile in Engineering Sciences (XE) (Top 15k candidates).
- GATE 2022: Attained 96 percentile in Mechanical Engineering (ME) (Top 80k candidates).