Rushil Manglik Email & Phone Number

San Francisco Bay Area

San Francisco, California, United States

Working with Dr. Cody Carroll

San Jose, California, United States

- Motivated product and feature prioritization - extracted a comprehensive product database from an unstructured corpus of 50 million product users, and improved accuracy through an NLP-based product categorization to 87%.- Developed a forecasting model leveraging sentiment scores extracted from news and social media modeled via OpenAI API and data from 150 million users as key features to improve the previous MAPE by 15%.- Collaborated with business and capability development team and deployed an end-to-end B2B knowledge sharing chatbot powered by LLMs (OpenAI).

Mountain View, California, United States

- Designed and implemented end-to-end deep learning pipelines to automate table detection and information extraction from documents, effectively eliminating information loss and minimizing manual transcription efforts.- Sped up model training time by 24x (48 hours to 2 hours) through hyperparameter optimization techniques using HPC.- Utilized Weights & Biases to log and track experiments, streamline model selection, identify the best performing model.- Improved Intersection-Over-Union metric from 0.4 (TableNet) to 0.97(fine-tuned model) for document table detection.- Utilized Amazon Textract's API to extract and convert textual and numerical information from tables into CSV and Pandas data frames, facilitating integration with TableQA NLP models for accurate question-answering.

Bengaluru, Karnataka

As the first and only Deep Learning Data Scientist in a life-sciences startup I achieved the following:Computer Vision:- Expanded services to biotech clients with cell engineering and screening studies, resulting in 3x revenue growth.- Built cloud-accessible deep learning pipelines to identify single cells encapsulated in water-in-oil nanodroplets by transfer learning using faster R-CNN resnet-50 model in PyTorch.- Achieved >97% accuracy on cell classification and localization, as against the baseline model's 50% accuracy.- Re-designed the in-house optical microscope, achieving uniform illumination and enhancing image signal-to-noise ratio and contrast, which enabled the model to accurately detect previously undetectable features.- Achieved 95% accuracy in extracting morphological features of localized cells by developing pipelines using OpenCV and scikit-image libraries in Python.- Visualized statistical distribution of cell count and its morphological features using Matplotlib and Seaborn in Python.Image Processing:- Contributed in a new version of AFM, leading to establishment of emerging marketing partnerships, and generating $100k worth of product sales.- Improved image resolution by 50% via deconvolution and removed image noise for Atomic Force Microscopy and Laser Scanning Microscopy images using scikit-image image processing library in Python.- Developed pipelines for AI-based software for detecting, measuring, and extracting nanoscale information such as cells, nano-particles, to reduce information loss and manual labor.

Bengaluru, Karnataka

-Tuned the PI controller in the Atomic Force Microscopy controlled loop and inferred that KP(proportional component) ranging from 10^-5- 10^4 and KI(integral component) ranging from 0.01 to 0.4 was the most acceptable value for the AFM in the AM(Amplitude Modulation) mode using the bode-plot model.-Conducted the AFM scans to demosntrate that the above values reduced the overall response time of the AFM controller loop by two third. This also reduced the artifacts due to various factors.-Image processing and analysis of the cells captured in nanodroplets using OpenCV library in python.

Kanpur Area, India

-Constructed a 2D electronic mesh phantom to mimic the electrical properties of a biological tissue(cervix), which will serve as a pioneer to build 3D electronic phantoms.-This will help improve our testing and analysis for early detection of cervical cancer by calibrating and assessing the EIT(Electrical Impedance Tomography) system and the reconstruction algorithm.-Implemented the forward solver for the EIT-based complete electrode model using finite element analysis and executed the numerical phantom in MATLAB.-The electronic phantom was built and discretized with the same number of triangular elements (3280) and nodes (1697) as the numerical phantom and was realized on the PCB using a resistor-capacitor network mesh designed on Altium software.-LabView subroutines (VIs) of the instruments used were designed and implemented for data acquisition and the data acquired was processed through a lock-in amplifier also designed and implemented on LabView.-The voltages and phases obtained from solving the forward model were compared with those obtained from conducting the experiments on the electronic phantom. The mean relative absolute errors between the two were within a 9% range.

Mumbai Area, India

Identified, fixed, and verified the bugs and loopholes in the logic developed by the Development Team for call routing. Achieved 100% success rate in production deployment.

Master Of Science - Ms, Data Science, 3.9/4.0

Courses: Applied Statistics, Data Acquisition, Linear Regression, Time Series Forecasting, Machine Learning, Deep Learning, Database.

Post Graduate Program, Artificial Intelligence And Machine Learning

Master Of Technology - Mtech, Electrical Engineering

Relevant coursework: Speech Signal Processing, Fuzzy Set, Logic and System and Applications, Management in Information Systems

Bachelor Of Technology - Btech, Electrical, Electronics And Communications Engineering

Activities and Societies: Public relations team lead at CULRAV-2016(annual cultural festival)Relevant coursework: Numerical Methods.