AI Researcher
Computer Vision • Federated Learning • Reasoning Systems
Integrated B.Tech + M.Tech student at IIT Kharagpur focused on building reliable learning systems. I work at the intersection of vision, representation learning, and federated settings, translating ideas into rigorous experiments and deployable pipelines.
I'm a Computer Science student at IIT Kharagpur with a focus on AI research, particularly in vision and federated learning. I enjoy building end-to-end research prototypes, from data curation and modeling to evaluation and reporting.
I've collaborated with IBM Research, KAIST, and Preimage on projects spanning reasoning data, representation learning, and 3D scene understanding. I value rigorous baselines, reproducibility, and clear scientific communication.
Pursuing integrated dual degree program in Computer Science and Engineering. Actively involved in research, competitive programming, and various technical competitions.
Built data curation workflow using FineMath, OpenWebMath, and Sangraha; automated Q&A pair generation for mathematical reasoning datasets; conducted empirical studies on test-time scaling behavior and performance trade-offs.
Devised pipeline leveraging Grounded SAM and MaskClustering for 3D segmentation on 360° images; integrated multi-view correspondence with vision-language models to track scene evolution and generate detailed progress reports.
Developed reward-prediction driven representation learning achieving up to 2x improvement in sample efficiency for RL tasks via LSTM-based temporal modeling; surveyed 40+ papers on action recognition and procedural understanding.
Integrated EfficientNet-V2-S with Adversarial Training on Purification (ATOP) to classify low-resolution real and AI-generated images. Achieved 43.77% F1 score for artifact identification.
Built a dual-agent dataset generation pipeline via Self-Instruct framework to create 2,000+ query-output pairs for instruction finetuning. Achieved 87.79% JSON similarity on tool-augmented LLM tasks.
IEEE ICDCSW 2025
Formulated a novel label-mixing framework for Vertical Federated Learning, reducing RMSE by up to 72% for regression tasks with minimal data overlap. This work addresses the challenge of data heterogeneity in federated learning scenarios where participants have different feature spaces.
I'm always interested in research collaborations, internships, and academic discussions. Feel free to reach out.