CV / Resume

Applied Scientist II, Devices & Services — Edge AI & Science

Amazon, Sunnyvale, California

• Proposed TL;DR, a reasoning-compression approach using thinking gradients and self-annealing rewards; reduced inference tokens by up to 32% on math and 20% on code while matching or exceeding baseline accuracy. [Under review]
• Proposed V-STAR, a GRPO-based post-training framework for grounded VLMs; reduced object-level hallucination by 23–44% and sentence-level hallucination by 60–73% across four architectures with no additional inference cost. [Under review]
• Co-led a study of how token compression and inference-time interventions compose in VLMs, achieving 89% token reduction and 54% hallucination reduction. [Under review]
• Developed video-aware SigLIP2 encoders that combat embedding collapse in static-camera retrieval, improving text-to-frame Recall@1 by up to 38.2 percentage points. [Under review]
• Delivered an invited tutorial on Efficient Edge VLMs at AMLC 2025 and shipped a multi-stage edge-VLM training pipeline adopted by multiple teams.

Applied Scientist Intern

Amazon, Sunnyvale, California & Seattle, Washington

• Designed attention-based embedding adaptation for sensor technologies, improving performance by 6% over the existing baseline.
• Built machine-learning models for early detection of AWS service failures, improving performance by 4% over the existing baseline.

Research Assistant

Machine Learning and Data Intensive Computing Lab, Rochester Institute of Technology

• Published 10+ papers, including first-author work on distributionally robust optimization, evidential learning, weak supervision, anomaly detection, and network sparsification.
• Developed methods for calibrated sparse networks, few-shot open-set recognition, and robust weakly supervised anomaly detection.

Ph.D. in Computing and Information Sciences

Rochester Institute of Technology, Rochester, New York

• Dissertation: Robust Weakly Supervised Learning for Real-World Anomaly Detection.

B.E. in Electronics and Communication Engineering

Tribhuvan University, Institute of Engineering, Pulchowk Campus, Lalitpur, Nepal

• Calibrated Knowledge Aggregation in Bayesian Mixture-of-Experts for Continual VQA. ICML.
• The Road Less Seen — Segment Exploration for Weakly Supervised Video Anomaly Detection. CVPR.
• Knowledge Exchange with Confidence — Cost-Effective LLM Integration for Reliable and Efficient Visual Question Answering. ICLR.

• GLEN — Generalized Focal Loss Ensemble of Low-Rank Networks for Calibrated Visual Question Answering. AAAI.

• Meta Evidential Transformer for Few-Shot Open-Set Recognition. ICML.
• Adaptive Important Region Selection with Reinforced Hierarchical Search for Dense Object Detection. NeurIPS.
• Reinforced Compressive Neural Architecture Search for Versatile Adversarial Robustness. KDD.

• Distributionally Robust Ensemble of Lottery Tickets Towards Calibrated Sparse Network Training. NeurIPS 2023.
• Adaptive Robust Evidential Optimization for Open Set Detection from Imbalanced Data. ICLR 2023.
• Balancing Bias and Variance for Active Weakly Supervised Learning. KDD 2022.
• Bayesian Nonparametric Submodular Video Partition for Robust Anomaly Detection. CVPR 2022.
• Distributionally Robust Optimization for Deep Kernel Multiple Instance Learning. AISTATS 2021.

• EB-1B Outstanding Researcher — U.S. permanent residency granted for outstanding contributions to machine learning.

• RIT Ph.D. Merit Scholarship — full financial support for doctoral studies.

• KDD Travel Award.

• Ncell Scholarship and Excellence Award.