Applied Researcher 4 at AI Foundations (LLM Core & Agentic AI), Capital One.
Previously,
Applied Scientist 2 at Microsoft
Ph.D. in Applied Mathematics from the University of California, Davis where I was advised by Dr. Thomas Strohmer.
My research focuses on Post-training Alignment of LLMs, Synthetic Data Generation, Differential Privacy, and Responsible AI.
Contacts:
Multimodal Learning
S. Broecker, J. Adams, G. Kumar et.al. 2025. “Multimodal Deep Learning for ARDS Detection”. Submitted, 2025.
Z. Huang et.al. “High-Fidelity Synthetic ECG Generation via Mel-Spectrogram Informed Diffusion Training”. GenAI4Health, NeurIPS, 2025.
Natural Language Processing and Synthetic Data
X. Yue et.al. “Synthetic Text Generation with Differential Privacy: A Simple and Practical Recipe”. ACL, 2023.
H. Shajari, G.Kumar et.al. “Short-text clustering with Large Language Models”. MSJAR, Microsoft Research, 2023.
C. Laterza et.al. US12105837B2 – “Generating private synthetic training data for training machine-learning models”. Patent, 2024.
D. Levitan et.al. US12353580B2 – “Building annotated models based on eyes-off data”. Patent, 2025.
D. Levitan et.al. US12517942B2 – “Dataset clustering and evaluation”. Patent, 2026.
Online Differential Privacy for Tabular and Spatial Data
G. Kumar, T. Strohmer, R. Vershynin. “An Algorithm for Streaming Differentially Private Data”. Accepted, Transactions on Data Privacy (TDP) 2026.
G. Kumar, T. Strohmer, R. Vershynin. “Continual Release of Differentially Private Synthetic Tabular Stream”. Submitted, 2025.
G. Kumar. “Differentially Private Synthetic Data Generation of Data Collected Over Time”. Doctoral Thesis, UC Davis, 2024.
Increased token efficiency of open-source models such as GPT-OSS 20B and Gemma 31B by 30-40%, while preserving accuracy on complex benchmarks (AIME, HumanEval) by utilizing Synthetic Data Generation (SDG) pipelines across teacher-student models to synthesize on-policy rollouts and contrastive preference pairs (ORPO/LCPO/SimPO).
Engineered an automated, multi-dimensional LLM-as-a-judge framework to evaluate, filter, and score millions of synthetic reasoning and tool-calling trajectories, eliminating evaluation bias and ensuring high-fidelity data for model post-training.
Architected scalable multi-turn data processing pipelines to ingest, validate, and standardize complex synthetic agentic data (including multi-step tool calls and execution feedback) into unified schemas for Supervised Fine-Tuning (SFT).
Scaled state-of-the-art LLM post-training pipelines across multi-node GPU clusters utilizing Megatron-Core, integrating Tensor (TP), Expert (EP), and Context Parallelism (CP) to train ~30B parameter architectures on synthetic datasets efficiently.
Designed a novel, LLM‑powered, agentic topic identification (text-clustering) system that provides 4x better quality topics compared to classical methods. Shipped as a skill in Excel Copilot. 2 Patents filed and an internal paper accepted.
Together with Microsoft AI Develop Acceleration Program (MAIDAP), delivered first ablation study on synthetic ECG generation via diffusion models conditioned on clinical notes (text) and patient demographics (tabular). Academic paper submitted.
Built a ML prediction model for NPS using 1,000+ sparse and noisy telemetry signals (activity, demographics, tenant data), identifying key drivers of customer SAT. Reduced noise by collaborating with domain experts on targeted user segmentation.
Built fairness into an LLM-powered feedback summarization tool, ensuring all customer segments (such as feedback source and sentiment) were accurately represented. Created an LLM-based evaluation framework for fairness.
Solved data retention and GDPR challenges by developing a differentially private synthetic text generation algorithm in collaboration with MS Research that fine-tuned a GPT-2 model on real data. Resulted in 2 filed patents, and 1 academic paper.
Shipped Differentially Private multi-document summarization models (BART, T5); contributed to the open-source microsoft/dp-transformers to enable T5 training with Opacus via custom PyTorch implementations of gradients.
Full-Stack Developer - Java, Angular, JavaScript, Camuda
Worked end-to-end on customer-focused applications as a Full Stack Developer using Java, JavaScript, Angular, and Camuda
Early on became a subject matter expert and was promoted from Analyst to Associate in only 2 years
Mentored 2 summer interns to develop a microservice for historic data maintenance using elastic search and Kafka
Awarded best student team in the United Nations (UN) Privacy Enhancing Technology (PET) lab's first-of-its-kind hackathon on eyes-off training of machine learning models on real user’s private data with differential privacy.
The scholarship is given to high-achieving mathematics undergrad or grad students of UC Davis.
Annual Internal Fellowship Competition for Continuing Graduate Students. Research funds being awarded by Graduate Studies and the Applied Mathematics Graduate Group at UC Davis.
Mathematical Foundations of Data Science, MAT-270, Spring 2022, UC Davis
Differential Calculus, Summer Session II 2020, UC Davis
Probability, MAT-135, Winter 2021, UC Davis
Applied Linear Algebra, MAT-167, Winter 2020, UC Davis
Algorithms Lab, Even Semester 2016, IIT BHU
Data Structures Lab, Odd Semester 2015, IIT BHU
Short Calculus, MAT-16A, Fall 2022, UC Davis
Linear Algebra, MAT-22A, Fall 2021, UC Davis
Calculus for Biology and Medicine, MAT-17A, Fall 2020, UC Davis
Multi Variable Calculus, MAT-17C, Spring 2020, UC Davis
Multi Variable Calculus, MAT-17C, Fall 2019, UC Davis