Computer Scientist | AI/ML Researcher
I have over 2 years of experience as a machine learning researcher and more than 2 years as a full-time software engineer in the industry. Currently, I work as an Applied Scientist at Johnson & Johnson MedTech, and our work aims to develop AI-driven solutions that enhance surgical precision, improve workflows, and lead to better patient outcomes. My interests lie at the intersection of machine learning and computer vision. I have developed expertise in applying machine learning techniques and utilizing ML frameworks to tackle complex problems in computer vision, such as high-accuracy object detection and tracking for autonomous driving, as well as deep learning applications in the surgical field.
Machine Learning Research Specialization
2022 - 2024
2016 - 2020
Here are some of the key skills that I've gained through my experiences:
Here are some of the key projects I've worked on as a Researcher at the COSMOS Lab, Columbia University.
Curated and released "Constellation", a dataset of 13K images suitable for research on detection of objects in dense urban streetscapes observed from high-elevation cameras. Benchmarked the latest object detection models on our dataset to observe performance on vehicle-pedestrian detection.
Trained and deployed deep learning video models for phase recogntion in Robotic-assisted Inguinal Hernia Surgeries. Research done in collaboration with Northwell Health / Lenox Hill Hospital Senior Surgeons and Medical Researchers, who provide the data and expert assessment of clinical performance.
Developed a computer vision system to automatically assess the skill level of robotic surgeons from videos of surgical trainees performing training exercises. The goal is to create a standardized system for robotic skill assessment that can be used to qualify surgeons nationwide. Research done in collaboration with Northwell Health / Lenox Hill Hospital Senior Surgeons and Medical Researchers (Paper coming soon.)
Other stuff that I've worked on in my free time.
Developed a multi-stage system combining object detection, perspective transformation, clustering, and graph neural networks (GNNs) to predict optimal soccer player passes in broadcast clips.
Achieved up to 7x speedup in inference time for Stable Diffusion v1.5 by combining several techniques including quantization, token merging, distilled UNet/VAE and JIT compilation.
Created an image captioning model for fashion data (DeepFashion Multimodal) using transformer networks to generate high-quality captions for finer attributes such as clothing fabric, jewelry, patterns, etc.
