May 2023 – Apr 2025 · 1 yr 11 mos · San Francisco, California, United States · Hybrid

• Worked on multimodal foundation models research and development, focusing on vision-language model architectures and training methodologies.
• Developed M3ID sampling method reducing VLM hallucinations by up to
• 28% without additional training. Published at CVPR 2024 and featured in AWS keynote at CVPR.

Aug 2020 – May 2023 · 2 yrs 9 mos · Sunnyvale, California, United States · Hybrid

• Halo Body
• Lead Research and Development to release Waist-Hip ratio and body circumferences as a part of Body scan on Amazon Halo. WHR is a better indicator of health risks than BMI and BFP.
• Trained an ML Model which is twice as accurate as state-of-the-art models for estimating body circumference.
• Our model is trained using only synthetic data but generalizes well when tested on real smartphone captured images.
• The results demonstrate that our system’s accuracy is considerably higher than that of self-reported manual measurements using tape measure and twice as accurate as other state-of-the-art models.
• Published in Nature Digital Medicine (2023) and filed multiple patents.
• Halo Trainer
• Built transformer-based models for fitness activity understanding, including on-device real-time
• repetition counting and form error detection. Won Best Paper Award at Amazon Computer Vision Conference 2023.

Oct 2017 – Aug 2020 · 2 yrs 10 mos · Sunnyvale, California

• 3D Object Detection: Lead Architect and Developer for 3D Object Recognition feature running in Cloud and deployed in ML-19. The pipeline was designed and built from scratch.
• Object Recognition algorithm is scalable in a large number of concurrent users, number of objects or size of map while maintaining a lower memory footprint.
• Led the team through research and development phase to deliver Object Recognition for ML19 OTA 1.
• Collaborated across other teams like Data, Cloud and Deep Learning teams to build evaluation and visualization tools.
• Object Recognition is a key feature for ML19 and one of the most requested feature for ML2 from developer feedback.
• Research paper is accepted at CVPR 2020 Workshop on AR/VR. Patent filed as well.
• Learned Keyframe Selection: Designed and implemented a PointNet based neural network which learns a frame embedding given sparse feature descriptors to improve retrieval for localization. Improved the localization recall with respect to Bag of Words algorithm by 10-30% while reducing the memory requirement to store each keyframe by 75%. Optimized the network resulting in similar computational requirement as Bag of Words.
• Scalable Infrastructure for SLAM Research: Designed and implemented a scalable infrastructure to extract data from various stages in the SLAM pipeline along with ground-truth. This enabled scalable training and evaluation of various SLAM related machine learning algorithms.

May 2016 – Aug 2016 · 3 mos · San Francisco Bay Area

• Research and development of algorithm to estimate the trajectory and stabilize loopy fyuses
• using factor graphs.

Aug 2012 – Aug 2017 · 5 yrs · Atlanta Metropolitan Area

• Distributed Object-based SLAM: Developed distributed algorithms using Distributed Gauss-Seidel methods for multi-robot trajectory estimation with minimal information exchange. Extended framework to include object-level semantics for distributed object-based SLAM. Published extensively in IJRR, ICRA, IROS with more than 500 citations.
• Memory-efficient SLAM: Proposed exactly sparse SLAM approach using multi-block Alternating Direction Method of Multipliers (ADMM) to enforce consistency among subgraphs.

May 2011 – Aug 2011 · 3 mos

• Developed a generic interface for all the search functions in Point Cloud Library
• Developed a fast Octree implementation on GPU.

Apr 2010 – Jul 2012 · 2 yrs 3 mos · Greater Hyderabad Area

• Bundle Adjustment on GPU: Developed hybrid CPU-GPU implementations of sparse bundle adjustment
• achieving 30-40x speedup over standard CPU implementations on datasets with up to 500 images using NVIDIA Tesla C2050 GPU. Published in ECCV 2010 workshop on computer vision on GPUs.
• Visibility Probability Structure from SfM Datasets: Developed visibility probability structures encoding
• visibility information between points and cameras as conditional probabilities for improved image localization. Published at ECCV 2012 with 86+ citations.

Jun 2009 – Jul 2010 · 1 yr 1 mo · Hyderabad, Telangana, India

• Research and development of medical image analysis algorithms for early detection of Diabetic Retinopathy