Sep 2022 – Apr 2024 · 1 yr 7 mos · Bengaluru, Karnataka, India

• Part of core research team of Amazon One. Helping scale Amazon One across the globe.
• Training and evaluation optimization:
• 1. Distributed training framework:
• Built distributed training framework to support multi-node multi-gpu distributed training over Sagemaker. The system supports using any kind of machine type (P3s, G5s)with GPUs while Sagemaker only allows P4/P3d machines. We achieved near linear speed improvement while scaling to multiple nodes depending on the model architecture. Visited Seattle to work in collaboration with software team for speeding up the process of delivery.
• 2. Improve training/evaluation speed and memory requirements by integrating deepspeed in the training pipeline, caching pre-processing outputs and improving data loading mechanisms to remove IO bottlenecks. Brief experience in model quantization and knowledge distillation for improving inference latency. Optimize inference on CPU by using frameworks like OpenVino and TVM without any loss of accuracy.
• 3.Workflow orchestration: Designed workflow orchestration using Managed Airflow to streamline training and evaluation of 100s of training jobs across team and also do scheduled training automatically with set conditions. Expected to reduce manual touch-points by 90% for all training and evaluations jobs

Sep 2021 – Aug 2022 · 11 mos · Bengaluru, Karnataka, India

• Part of core research team of Amazon One(one.amazon.com). First MLE hire in the research team.
• Root-cause automation workflow:
• ▪️ Formulated the problem, designed and delivered RCAEngine, a multistage system for automatic root cause analysis of biometric images flagged by the live system, replacing a manual process and saving approximately 50 person hours per week. This also streamlined the process of root-causing while removing subjective judgement of different persons when looking at images. Lead to expansion of the product into states with stricter data retention regulations and reduction in the biometric customer data retention policy from 7 days to 3 days.
• ▪️ Developed a cascaded rule engine using signals from multiple vision-based models and custom image analysis to accurately classify root causes at 98% and provide insights for system generalization.
• ▪️ Implemented CI/CD pipeline for RCAEngine, improving the efficiency and reliability of the deployment process, and successfully migrated existing infrastructure as code packages written with CloudFormation (internally deprecated) to a compute environment using AWS Cloud Development Kit (CDK), resulting in scalable and flexible system.
• ▪️ Spearheaded the setup of a data pipeline that resulted in the creation of self-service and customizable Tableau dashboards for performance reporting used by senior leadership and business teams, reducing the ops load for the research team by approximately 80% and enabling more efficient decision-making through real-time visibility into key performance metrics.
• ▪️ Collaborated with cross-functional, cross-timezone teams to successfully deliver the first project in the newly formed India research team within 1 year, earning trust and establishing a strong track record of successful project delivery.

Jan 2021 – Aug 2021 · 7 mos

• 1) Video Analytics Framework: Designed and developed a complete scalable framework for easy integration of video analytics solutions as a service in this framework. This framework is used by our team to deploy any video analytics solutions. The current services include face recognition at scale(Edge compatible), Number plate recognition system(Edge compatible) , Equipment monitoring in a factory based setup.
• Connections of backend and frontend is WebSocket based, kibana(elasticsearch) is used for dashboarding, Mongo for database, AWS for cloud service.
• 2)Smart powerpoint redaction system : Reduced 50% person hours for one of the largest consulting firms in the world by automatically redacting sensitive content(logos, revenue figures, charts, client names, etc) from powerpoint presentations to be able to adhere to privacy and security policy while saving powerpoints by building NER models for textual content, object detection models for images and statistical manipulations for redacting charts. Designed a serverless architecture for productionizing the complete system on AWS.

Jul 2019 – Dec 2020 · 1 yr 5 mos

• 1) Information Extraction
• > Document information extraction: Reduced 70% manual labour for a worldwide automotive leader by developing modelsfor table information detection/extraction and automatic key-value pairs extraction. Used efficientDet for table detection,pix2pix GANs and image processing for table structure extraction and cloud OCR service to recognize text. Designedcomplete architecture on AWS Cloud using Lambdas, EC2 and S3.
• > Automatic Information Extraction From Shipping Labels: Reduced 90% manual labor by automatically extracting name and address components from images of the delivery invoices of multiple vendors such as Amazon, FedEx, Ups, etc. Deployed the system on Ipad pro using coreML framework. The POC was done on Samsung galaxy Tab 2019 edition.
• 2) Natural language Processing
• >Question Answering System: Implemented an end to end question answering system for a huge corpus of pdf manuals.
• >Food search using AI: Improved the conversion rate from user search to actual order by 40% for a digital ordering startupbased in California. Used custom embeddings, siamese network with triplet loss and elasticsearch to improve the search results for their food ordering platform.
• 3) Miscellaneous
• >Image Search: Built an Image Search System that finds images that are most similar to a query image from a large dataset of images, thus making better recommendations for a user. The model was implemented using the deep ranking technique with more than 60% recall across all categories.

Jul 2018 – Jun 2019 · 11 mos

• > Resume ranking: Transformer based Model that extracts all the relevant information in the resume like person's education, company and personal details and uses this information to find similarity and ranks them based on the job description. We were able to achieve an accuracy of 88%. This project helped the human resource team to shortlist resumes based on job description apart from that it helped the TAP team to automate the process of project allocation based on persons's skill set and work experience.
• >Table Detection and Automatic Information Extraction from images of Tables with limited scope using GANs.

Jan 2018 – Jun 2018 · 5 mos

• First Trainee hired in the team.
• >Exploratory Data Analysis: Extensive Exploratory Data Analysis on a multi-level dataset. Built base classification models to predict intermediate features.
• > Wrote library to randomly generate chart data with object detection annotations using Matplotlib and Seaborn to completely automate data generation for any type of chart data with annotations.
• > Line Chart Knowledge Extraction: Line Chart Knowledge Extraction is a tool that automatically extracts information from research paper line graph images with a limited scope.

Dec 2019 – Jan 2020 · 1 mo

• Link: https://drive.google.com/open?id=1SNs5zsQb6P2Tg-yqrvOEDJvdzrfPeGvq

May 2017 – Jul 2017 · 2 mos · Noida Area, India

• Remote Video Surveillance: Designed and Implemented the back end for a RVS System using Python, FFMPEG and openCV. Deployed the system on AWS and completely automated AWS management steps using boto library of python.
• Data Augmentation: Generated different perspectives of an image using the keras library ofpython and produced over 10 million images from 0.6 million images.
• Made frame extraction from video 50% more efficient by splitting videos in segments and mapping the frame numbers from the large video to the splitted videos.
• Explored basics of Neural Networks and Regression Analysis