cv

Summer Research Assistant

Johns Hopkins University - Center for Language and Speech Processing

Supervisors: Prof. Daniel Khashabi, Prof. Eric Nalisnick

• Building a PyTorch and vLLM-based inference engine for budget-aware decoding on mathematical reasoning benchmarks, maintaining performance degradation within 5%

Research Assistant

Carnegie Mellon University - InfiniAI Lab

Supervisor: Prof. Beidi Chen

• Optimizing speculative decoding, Key-Value (KV) cache compression, and continuous batching within MagicDec, achieving up to 2.5× throughput for long-context LLMs (32K tokens, batch size 256) on A100 GPUs

Graduate Research Assistant

Johns Hopkins University - ARCADE Laboratory

Supervisor: Prof. Mathias Unberath

• Built a language-promptable digital twin using a multi-modal foundation model (FluoroSAM) for real-time segmentation, 3D reconstruction, and automatic collimation, reducing radiation exposure by 60%
• Engineered an LLM-driven interface to interpret complex verbal instructions for the Brainlab Loop-X robotic C-arm, allowing autonomous, hands-free surgical imaging with 84% end-to-end success in cadaveric trials

Senior Engineer - Machine Learning

Bosch Global Software Technologies

• Spearheaded a team of 6 to develop an AI-powered test case optimizer for the Automotive Electronics Division, reducing test runtime by 66% and increasing code coverage to 94%
• Delivered scalable chatbot automation for the Bosch Automation Platform, integrating Named Entity Recognition and ElasticSearch to improve search relevance, boosting customer satisfaction by 31% across 5,000+ enterprise users

Summer Research Fellow

University of Zürich - Artificial Intelligence and Machine Learning Group

Supervisor: Prof. Manuel Günther

• Boosted face recognition on low-quality surveillance footage by designing a Super-Resolution GAN, increasing accuracy by 37% on downsampled and 5% on full datasets for reliable identification in real-world security scenarios

Master of Science and Engineering

Johns Hopkins University

Computer Science

Bachelor of Technology

National Institute of Technology, Tiruchirappalli

Electronics and Communication Engineering

Selective KV Quantization

Don't Drop It, Compress It - Reducing LLM KV cache memory usage

• Led a team of 4 to design and implement Selective KV Quantization for LLM inference, preserving sink/window tokens in full precision while quantizing older cache entries to int8, achieving 2× memory savings with minimal impact on perplexity (5.56) and ROUGE-L (0.2073 → 0.1709)

Rich Teacher Features for Efficient Single-Image Haze Removal

Lightweight haze removal pipeline via knowledge distillation

• Devised a lightweight haze removal pipeline via heterogeneous knowledge distillation with a novel feature affinity module, yielding a 15% PSNR gain and 20× model compression

Intelligent Control of Robotic X-ray Devices using a Language-Promptable Digital Twin

IPCAI

Rich feature distillation with feature affinity module for efficient image dehazing

Optik Journal

The Siemens Healthineers Best Paper Award

IPCAI 2025

Best Project Award

NLP: Self-Supervised (Spring 2025), Johns Hopkins University