About Me

I’m working as an AI Researcher at Together AI, on pre-training and finetuning Large Language Models, cooking up data recipes, and building out training framework features as needed. Lately I’ve been digging into long-context training and model distillation.

Before this, I got my MSc in Computer Science from NYU’s Courant Institute of Mathematical Sciences, where I explored Reinforcement Learning from Human Feedback for LLMs at OLAB, NYU Langone. I’ve also done research at CVIT Lab at IIIT-Hyderabad along with industry data science roles, working on information extraction, machine translation, and question-answering.

These days I’m most excited about RL for language models, systems work around attention, and multimodality. Also I keep an eye out for ML infrastructure.

Check out my CV for more detailed info.

Featured Projects

Explored retrieval strategies for RAG pipelines including chunking optimization, sparse vs. dense embeddings (TF-IDF, DRAGON+, Sentence Transformers), and two-level retrieve-and-rerank approaches. Achieved 78.6% Recall@10 through rechunking and similarity thresholding.

Retrieval Optimization

LLMs in healthcare produce untrustworthy outputs due to biases from human annotators in RLHF. We leverage structural bias in multiple-choice questions to generate alignment data at medical licensing exam complexity, achieving +5.5% over SFT on MMLU medical subsets.

DPO-RLIF: Reinforcement Learning from Structurally-Derived Implicit Feedback

Mixture-of-Experts models suffer from inefficient batched computation due to dynamic routing and load imbalance. We developed variable-sized batched GEMM kernels using cuBLAS as an alternative to MegaBlocks' Triton-based sparse kernels. Outperforms sparse MegaBlocks and torch.bmm.

Exploring cuBLAS-Based Variable-Sized Batched GEMM for MoE Inference

Paper Digest

FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning February 01, 2024

Visual Instruction Tuning (LLaVA) January 20, 2024

DINOv2: Learning Robust Visual Features without Supervision March 01, 2024

View all →

Selected Talks

GPT-3: Language Models are Few-shot Learners - Large Language and Vision Models Symposium, NYU Center for Data Science, 2024

Llama 2: Open Foundation and Fine-Tuned Chat Models - Large Language and Vision Models Symposium, NYU Center for Data Science, 2024

Generative Agents: Interactive Simulacra of Human Behavior - Large Language and Vision Models Symposium, NYU Center for Data Science, 2024

View all →

Conferences

MLSys 2025 - Conference on Machine Learning and Systems

Conference takeaways

NeurIPS 2025 - Conference on Neural Information Processing Systems

View all →