Red Hat Inc.
Senior Machine Learning Research Engineer
Oct 2025 – Sep 2026
Machine Learning Research team, Red Hat AI
Post-Training & Inference Optimization
- Drove research on post-training and inference, including mixed-precision quantization, parallel token drafting for speculative decoding, and RL rollouts
- Integrated FP8 quantized Inkling, Qwen 3.6, Granite4 and Nemotron family of models into vLLM and llm-compressor
- Developed Dflash drafters (parallel token predictors) for Qwen3 and Nemotron
- Mentored one intern on the KV-Dflash project
Open-Source Contributions
- Contributed to open-source frameworks: vLLM, llm-compressor, speculators and the Red Hat AI Hugging Face model repository
- Enabled production deployment of compressed and accelerated LLMs
Agentic AI & Evaluation
- Worked on Agentic AI evaluation, including SWE-Bench
- Built and curated tool calling and Agentic datasets, along with their evaluation pipelines
Argonne National Laboratory
Postdoctoral Researcher
Aug 2023 – Sep 2025
Supervisors: Murali Emani, Venkatram Vishwanath
ImageNet-Think-250K
- Designed a synthetic multimodal reasoning dataset with 500K structured thinking–answer pairs to enable training and fine-grained evaluation of explicit visual reasoning
- Synthetic data generation captures step-by-step Chain-of-Thought style reasoning traces from multiple thinking VLMs
- Established comprehensive evaluations by benchmarking several contemporary reasoning-centric VLMs across multiple metrics, providing a standardized testbed for multimodal reasoning quality and interpretability
PagedEviction
- Developed a structured, block-wise and attention-agnostic KV cache eviction algorithm, specifically designed for vLLM's PagedAttention
- Requires minimal changes to vLLM's core KV Cache implementation and improves LLM inference on long context benchmarks
- Achieves similar performance on LongBench datasets compared to existing methods, while boosting inference performance
LExI
- Designed a novel, data-free post-training method to choose the optimal number of active experts per layer in MoE models
- Improves inference performance and accuracy over MoE pruning techniques on several MoE models such as DeepSeekV2, OLMoE-1B-7B, Qwen-1.5-MoE, Mixtral
- Achieves up to 10% higher accuracy and 2x better vLLM throughput
MoPEQ
- Developed a novel mixed precision quantization algorithm for MoE-based Large Language and Vision Models, enabling per-expert adaptive bit-width assignment (2, 3, 4 bits) using Hessian-based sensitivity of each expert
- Achieves memory reduction while maintaining accuracy on DeepSeek-VL2 tiny, small, base, and MolmoE models
- Outperforms baselines on several VLMEvalKit datasets
LangVision-LoRA-NAS
- Designed a framework to integrate NAS with LoRA to optimize the finetuning efficiency of LLMs/VLMs
- Leverages NAS to dynamically identify optimal LoRA ranks across different layers
- On LLaMA-3.2-11B-Vision model across diverse vision-text datasets, demonstrated significant improvement in LoRA trainable parameters while preserving the baseline perplexity
LLM-Inference-Bench
- Developed a comprehensive benchmarking and profiling suite for LLM inference (LLaMA-2/3, Mistral, Qwen) on diverse accelerators (NVIDIA A100/H100/GH200, AMD MI250/MI300X, SambaNova SN40L, Habana Gaudi2)
- Evaluated using vLLM, TensorRT-LLM, llama.cpp, and DeepSpeed-MII frameworks
- Provides crucial insights into the scalability and efficiency of LLMs by examining the interplay between input length, output length, and batch size with various inference optimization techniques
- Robust infrastructure used for Argonne's AuroraGPT project
WActiGrad LLM Pruning
- Developed a structured LLM pruning method to prune attention heads, MLP dimension and hidden size across different layers of LLaMA and Mistral models
- Integrated the pruned LLMs into different inference frameworks and achieved enhanced performance on Nvidia A100 GPU, Cerebras CS-2, Groq LPU and Graphcore Bow Pod64 accelerators
- Achieved similar perplexity on WikiText, C4 and PTB datasets
Iowa State University
Graduate Research Assistant
Aug 2017 – Aug 2023
Supervisor: Dr. Arun K. Somani
Accelerator, Architecture and Mixed Precision Quantization Co-Search
- Developed Fast Differentiable Hardware-aware Mixed Precision Quantization Search for optimal precision per weight and activation matrix
- Joint Differentiable Architecture and Mixed Precision Quantization Co-search for kernel/filter size and precision
- Joint Accelerator, Architecture, and Precision triple co-search for optimal hardware-model combinations
- Demonstrated on Bitfusion accelerator with MobileNetV2, achieving better accuracy-latency trade-offs than baselines
ConVision Benchmark
- Developed a comprehensive PyTorch framework to standardize implementation and evaluation of CNN and ViT models
- Implemented 14 CNN model families: AlexNet, ConvNext, DenseNet, EfficientNet, GhostNet, Inception, MNASNet, MobileNet, NFNet, RegNet, ResNet, ShuffleNet, SqueezeNet, VGG
- Implemented 26 ViT model families including BoTNet, CCT, CaiT, CrossFormer, Swin, MaxViT, and more
- Rigorous evaluation: accuracy, precision, recall, F1 score, FLOPs, MACs, CPU/GPU latency
Array Aware Neural Architecture Search
- Designed NAS to automatically produce efficient CNNs for fixed-size systolic-array accelerators
- Novel CNN search space based on underlying hardware array dimensions for optimal performance
- On CIFAR-10, achieved similar accuracy as baselines while saving substantial accelerator cycles
Hardware Dimension Aware Pruning (HDAP)
- Developed HDAP for systolic arrays, multi-core CPUs, and Tensor Core GPUs considering hardware dimensions
- Prunes nodes/filters per layer based on underlying hardware size in every iteration
- 3.2x speedup on Turing Tensor Cores (baseline: 1.5x) and 4.2x on Eyeriss accelerator (baseline: 1.6x)
Fault and Array Size Based Pruning (FPAP)
- Co-design method to bypass faults and remove internal redundancy concurrently for efficient inference
- Compared with different pruning methods under random, row, and column fault scenarios
- 4.2x mean speedup (baselines: 1.6x) on ConvNet, NiN, AlexNet, VGG16 over Eyeriss with random faults
Argonne National Laboratory
Research Intern
Sep 2021 – Nov 2021
Supervisors: Murali Emani, Venkatram Vishwanath
Searching Sparse and Mixed Precision Quantized Neural Networks for A100 Tensor Cores
- Developed Mixed Sparse and Precision Search (MSPS) for optimal weight matrix type (sparse/dense) and precision per layer
- MSPS outperformed manually designed Int8 ResNet50 in accuracy and latency
- Extended to Architecture, Sparsity, and Precision Search (ASPS) for joint hyperparameter optimization
- Best ASPS model: 1.1x faster and 0.57% more accurate than baseline sparse Int8 ResNet50
Intel Corporation
Deep Learning Research Intern
Jun 2020 – Dec 2020
Supervisor: Sreeni Kothandaraman
Searching Architecture and Precision for U-net Based Image Restoration
- Weight sharing NAS to optimize U-net for Super Resolution and Denoising
- Operation Search for optimal convolution and pooling sequences in Down/Up sampling
- Experiments on DIV2K and BSD400 datasets for Super Resolution and Denoising
- Searched U-net outperformed baseline; mixed precision achieved better PSNR than uniform quantization
Advanced Micro Devices (AMD)
Deep Learning Intern — MIGraphX Team
May 2019 – Aug 2019
Supervisor: Mike Vermeulen
Post Training Quantization for CNN Models
- Developed PTQ methods to reduce CNN weights from FP32 to Int8 precision
- Implemented and analyzed on VGG16, ResNet50, InceptionV3, Xception with negligible ImageNet accuracy loss