Projects

O3SLM: Open Weight, Open Data, and Open Vocabulary Sketch-Language Model

Published in 40th AAAI Conference on Artifical Intelligence, 2026

Tackled the gap in LVLMs’ ability to understand hand-drawn sketches by introducing the first large-scale dataset linking sketches, real images, and natural-language instructions, and an LVLM trained on it. Across tasks such as localization, counting, retrieval, and VQA, our model achieves state-of-the-art performance, significantly advancing sketch-based visual reasoning.

Download Paper | Download Bibtex

Sound Classification in Indian Cities

Published in Tiny Papers Track at ICLR, 2024

Developed a multi-label urban sound classification model using transfer learning on a manually curated dataset from Indian cities. Fine-tuned VGGish achieved a 32% F1-score improvement over AudioSet-trained baselines.

Download Paper | Download Bibtex

Conditional LiDAR Point Cloud Generation using Diffusion Models

This project explores diffusion models for generating LiDAR point clouds with conditional control over object positions, enabling scene manipulation and data augmentation. It is the first work to achieve controllable object manipulation in diffusion-based LiDAR generation

Few-Shot Learning for Intent Detection

Developed a few-shot episodic learning pipeline for intent detection on the banking77 and clinc150 datasets on PyTorch. Implemented a cross-encoder and parametric network for similarity scoring.

Pretrained Language Models for Molecular Activity Prediction

End-to-end finetuned various language models for predicting various biochemical properties of biomolecules. I sourced data from the CHeMBL database to predict the drug likeness of a biomolecule.