Srinidhi M Email & Phone Number

San Francisco, Ca, Us

San Francisco, California, Us

Seattle, Wa, Us

Neo AI Compiler team• Designed and implemented support for Sagemaker scikit learn extension to include the new scikit-learn estimators. Contributed to the parser (relay IR) and the compiler stack (TVM) • Designed and led the software release process for Sagemaker Edge. This entailed triggering long-haul tests securely on device farm, reporting the results, packaging the software and reliably releasing it to the customers • Contributed to a project to enable TensorFlow’s Quantized Neural Networks to work with TF-TRT along with TVM• Contributed to the TVM stack under various projects to support new deep learning models

Seattle, Wa, Us

Amazon Physical Stores - "Just Walk Out" Technology• Worked closely with researchers at Amazon Go. Optimized the models on the training platform using methods involving profiling and optimizing numpy/gpu calls which increased per epoch training speed by ~5x and eval time by ~3x. Improved data ingestion by exploiting shared name prefix in aws s3 and multithreaded s3 calls. This resulted in major improvements such as bringing down training time from 60 hrs to ~16 hrs and data loading time from ~10 hrs to ~1 min• Re-designed and implemented a categorized cost tracking of model training to increase the transparency with usage reports and tagged metering for each model/team. This work ensured backward compatibility to org wide EC2 billing

San Jose, Ca, Us

• Part of Application Policy Infrastructure Controller (APIC) which is the main component of Cisco's Application Centric Infrastructure (ACI) group that provides software-defined networking (SDN) solutions. The controller manages and operates a scalable multi-tenant Cisco ACI fabric.• Building a Natural Language Processing (NLP) assistant - for APIC on which users can raise their network related issues and query information on their network nodes as a text query• Designed a system to parse a text query intelligently in several NLP stages such as implementing sentence segmentation, shortest dependency path, logistic regression and biLSTMs with attention, which achieved near-human level performance in parsing them to a formal language required for searching our Distributed Graph Database. Created, cleaned and processed data required to train the NLP models. Currently, working on incorporating text autocompletion and suggestions.• Implemented kafka consumers and producers using shopify/sarama library in golang to efficiently handle the logs of user-induced changes in the network on multiple machines• Proposed and executed a system with docker images and API calls that can efficiently and concurrently communicate with several other elements of our project

New York, Ny, Us

• Graduate Teaching assistant at NYU for the course Social Networks (CSCI-UA:0480-004) for Fall (2017) and part-time adjunct Professor for the course PACII (Intensive Fast track EECS course) for Spring (2018)), OS and Fundamental Algorithms for fall (2018). Delivered weekly recitations, held office hours for 40+ students and graded papers/HWs for students

• Implemented various objection detection and instance segmentation algorithms such as Faster RCNN, Mask-RCNN etc, for pneumothorax and chest tubes detection in chest MRIs, which were used as suggested annotations for radiologists• Wrote and presented an abstract on "Crowdsourcing pneumothorax annotations using machine learning annotations (MLA) on the NIH chest x-ray dataset" which was selected for Annual SIIM Scientific Conference on Machine Intelligence in Medical Imaging (C-MIMI). • Developed the infrastructures for hyperparameter tuning, model training and deployment using Amazon Sagemaker, ECR, S3 and Docker containers• Built a general-purpose segmentation DL network for 'Medical-Decathlon' competition, aiming to generalize over multiple segmentation tasks with different modalities such as Liver/tumor from CT scans, active tumor/oedema from Brain tumors in MRIs • Detected and classified lung nodules using Deep lung - an architecture used to segment lungs and nodules on them, from MRIs

Kharagpur, West Bengal, In

Title: "Closed-Loop Low Powered Electrode Stimulator and Recorder"Research Project as part of Bachelor's Degree Thesis under Prof. Sudip Nag(E&ECE Dept.) at IIT Kharagpur• The ability to record and to control action potential firing in neuronal circuits is critical to understanding various neural functions, automatic stimulation of muscles, understand the functions of brain, locomotion of spinal cord injured humans, cardiac pacemaker and others. This project explores the system level design of closed-loop stimulation and recording using both low power stimulator and an improved low power amplifier. The objective of this study is to develop a monolithic integrated circuit (IC) to record action potentials and simultaneously control action potential firing with an aim to run embedded neural networks on it. Both these circuits were merged effectively using divided power sources, thus sourcing both the analog and digital part of the circuits. • The report attached explains design challenges faced, the architectures implemented, the tool set this project provides to achieve various functionalities and their reasons while simultaneously implementing a state-of-the-art low power architectures. • Finally the results have been summarized showing a completely designed PCB board.

Suwon-Si, Gyeonggi-Do, Kr

Project: "White paper for 'SSDs over PCIe over NVMe over Ethernet' "• Part of the few from the beginning of the project. Successfully submitted the research report on the titled topic that possibly directs new research for the corresponding research team in Samsung Electronics Research, India• Part of my job included presenting my work every week on white board • Studied basic technologies and protocols such as HDDs, SSDs, PCIe, NVMe , SATA, NVMe over fabrics, NVMe over Ethernet, NVMe over InfiniBand, RDMA, ROCe vs iWARP and others. • Made a report stating advantages and disadvantages of each available technologies mainly focusing on PCIe SSDs over NVMe interface protocol over Ethernet networking protocol, about the companies and research groups working on each existing technology and aiming for the state-of-the-art future technologies, the new and most efficient devices for each technology in the market and other new architectures proposed in events such as OCP(Open Compute Project) Summits• Presented my own insights on where the technology must be improved and where it lacks research• In addition, personally requested for a project and was assigned to write a Verilog code for tournament event scheduling, tested it and it is presently being incorporated into other modules at the company• Received a pre-placement offer to work there

Project-"Restoring Remote Sensing images of invariant and variant blur"Supervisor : Dr. R. Chandrakanth (Head of Image Processing Div.)• Gaussian blur kernels with constant and varying deviation(sigma) were used to create both spatially variant and invariant images• Implemented various techniques for deblurring from a single image such as Inverse Filter, Weiner Filter, Wavelet Restoration, Blind Deconvolution, Landweber iteration and Richardson-Lucy Algorithm and finally compared the results• Further explored various algorithms such as one proposed by Cheong et al in "Fast Image Restoration for Spatially Varying Defocus Blur of Imaging Sensor" and others, worked towards increasing the robustness of their algorithms

Master'S Degree, Computer Science

Bachelor'S Degree, Electronics And Electrical Communication Engineering(Major), Computer Science(Minor)