Optimizing clinical operations and care with a unified platform powered by finely tuned LLMs.

I am a tech lead primarily specializing in solving perception problems using deep learning based techniques. I have worked on image segmentation, 2D/3D object detection and depth estimation using camera, lidar and radar sensors along with approaches to fuse the sensor data. I have played major role in bootstrapping ML infra, ML frameworks and using Simulation data for ML experiments at Cruise. I love exploring the breadth of problem space and doing whatever it takes to solve challenging problems. Some of the problems that I solved include:- Building and deploying an extremely efficient and accurate road detection system within the first month of joining- Built ML framework based on Tensorflow that enables rapid model development and iteration. Every model in Cruise was built and deployed using this- Lead a team to build ML infra to enable large scale training (up to 512 gpus) of ML models and optimizing them further thru hyper-param tuning- Built first of a kind prototype of camera+radar fusion model for 3D object detection and segmentation- Lead a team to train and evaluate perception models using realistic simulation dataset including setting up a data pipeline to generate large scale simulation data. This worked well enough for the company to bet on it for all future tech development and validation- Lead a team to build a pipeline and associated tooling for statistically measuring simulation fidelity across all sensor types and different simulation componentsApart from tech development, I also helped lead several other initiatives like Cruise car deployment in Japan, ML platform development, onboarding new engineers and upskilling existing engineers through tech seminars and workshops.

• Architected and built a software platform that enables one to design, deploy, manage and monitor ML pipelines intended to keep user accounts safe from abuse and compromise. This powers some of the largest ML pipelines at Microsoft processing 10s of TB per day, running on 100 thousand nodes and affecting billions of user interactions. Improved compute performance by up to 40% and reduced data storage needs by 30% by implementing cutting-edge big-data performance optimization algorithm.• Re-architected and migrated an existing distributed web service to Azure resulting in 80% reduction in compute/storage usage. Simplified 17 different services into one simple service while improving code quality and speed from two months for a single line of change to one click deployments every week with best in class monitoring.• Built a prototype version of ML system that is currently used in serving Microsoft Captcha• Prototyped a behavioral telemetry service that collects metadata from millions of users in order to detect and prevent abuse.

● Researched and developed a tool to auto-detect a credit card like object in a sequence of images which is used to estimate scale of objects in the image.● Implemented an image enhancement tool to automatically improve the visual appeal of images by reducing noise and improving contrast and white balance.● Implemented a robust skin detector using naive Bayes classifier on LUV color space.● Prototyped a background segmentation engine capable of identifying user in the foreground and masking the background using active contours.● Implemented a Flash based application for guiding users through a step-by-step process to create a virtual avatar using Viola-Jones face detector and motion tracking algorithms.● Developed a viral Facebook application that detects facial features on user’s profile picture, using Adaboost classifier trained on LBP feature descriptors, and applies a mask.

● Designed and implemented a web service for accessing enterprise document repositories with support for third-party authentication modules.● Implemented a Kindle application for providing access to document repositories for enterprise customers.● Prototyped a software component that incorporates Facebook friend recommendations into Amazon product view.

● Developed a software component in Jugular Graphics Engine to partition complex 3D tessellated geometry into simpler primitives using Principal Component Analysis.