Machine Data

Reddy Vishnuvardhan Reddy

Email: ********@****.***,*****************@*****.***

Linked in: https://www.linkedin.com/in/reddy-vishnuvardhan-challapalli-bb1b99110/ Website: https://vishnuvardhan1028.wixsite.com/website Git hub: https://github.com/vishnuvardhanreddychallapalli SKILLS

PROGRAMMING:

Comfortable:

Python• C • C++• Java• STL(C++)

•HTML•CSS•JavaScript•NodeJS

Advanced Learner:

Jupyter Notebooks • Pandas • Numpy

• Scikit Learn

Intermediate:

TensorFlow • Kera’s • Spark •

Hadoop•Pytorch•OpenCV•A/B Testing

EDUCATION

University of North Carolina

M.S COMPUTER SCIENCE

Fall 2017 to December spring 2018.

Grade Point Average:3.5/4

OUTSIDE WORK

Algorithms implemented

from scratch in python and

C++:

•Decision Trees

•K-NN(locality sensitive Hashing)

•Naïve Bayes (Multinomial)

• Linear and Logistic Regression.

COURSEWORK

MOOCS

•Deep Learning by Stanford University

•Build Self Driving Cars by MIT

University of

North Carolina

at Charlotte:

•Machine Learning

•Algorithms

•Data Mining

•Intelligent Systems

•Parallel Computing

•Cloud computing

•Database Systems

Achievements

•top 2% in the class of Bachelors

Degree.

•Two times Hackathon winner in

Anna University Digital Design

Competition using VHDL

EXPERIENCE

Deka Research and Development Machine learning Software Engineer

February 2019 Manchester New Hampshire

• Worked on applying machine learning techniques and Sensor Fusion Approaches to build Machine Learning models using Robot Operating System (ROS).

• Deployment Object Detection algorithms like SSD and Faster RCNN in C++ using NVIDIA Tensor RT.

• Implemented Free Space Estimation that Generates Occupancy Grid using Ultrasonic sensors that runs at 5 frames per second using Open MP in c++.

•Implemented Hungarian Algorithm for 2X Faster data association in Object Tracking.

•Worked on Semantic Segmentation with focal loss and improved the IOU metric by 10% and Used Knowledge Distillation to bring the model size by 5 times smaller.

•Deployed Deep Neural Network Architectures for Semantic Segmentation using Tensor RT NVIDIA framework to achieve 30Hz frequency.

•Worked with a Proof of Concept for Car Blinker Detection using Conv-LSTM and Encoder Decoder Models with Attention Mechanism.

•Implemented Cuda kernels for SoftMax and Transpose Convolutions using CUDA on GTX1060-T that runs 30 times that runs faster than CPU Parallel Algorithms.

PROJECTS

Implemented a Self-Driving car model using Deep Learning with Object Detection

USING CONVOLUTIONAL NEURAL NETWORKS PYTHON and

TENSORFLOW and OPENCV

• Implemented the NVIDIA Self-Driving neural network using Keras to train a video streaming recorded from the front dash cam of the car.

• Framed the problem to predict the steering angle of the car given an image of the car by converting a video to sequence of images.

• Observed the accuracy to be 93% when I applied the neural architecture similar in the paper.

•Improved the accuracy of the model by adding drop outs and Batch Normalization techniques and proper weight initializations like Xavier-glorot techniques to 98%.

•Applied Tensor Flow object detection model with Faster R-CNN on the video to detect cars in the video.

Recommendation Systems on Movie dataset Tensor

flow and scikit learn and spark MLLIB Spark Graph X

•This is Graph based data used spark Graph X module to process this data.

•Used several traversal techniques like Breadth first search and depth first search to store relative meta information by graph traversing for computing similarities between movies and users.

•Applied Classical Machine learning techniques like collaborative filtering techniques like item-item similarity and user-user similarity and Singular Value

Decompositions or Matrix

Factorization techniques.

•Applied Machine learning

techniques like Gradient

boosting decision trees on the

dataset and observed a

dramatic improvement in the

model by 2 times as that of

Collaborative techniques to

see that content based

recommendations works

better than collaborative

filtering techniques.

•Tried both LSTM and

Gradient recurrent units and

found GRU’s work better than

LSTM.

Amazon Fine Food

reviews(Given a text find

the polarity of the review)

Tensor flow and scikit learn

•Applied Data cleaning

techniques like Bag of words,

TF-IDF, Word2vec

•Used naïve bayes and

Gradient boosting decision

trees but accuracies are not

higher because of the

improper word embeddings

•Applied LSTM’s on the

dataset but trained word

embeddings with respect to

data instead of using

word2Vec and Glove

Techniques observed an

accuracy of 95%for 3 epochs.

Implemented the

following algorithms in

Mapreduce using Spark

without MLLIB Used Scikit

learn and Numpy and pandas

•K-means

•K-nearest Neighbors

•Logistic Regression and

Support Vector machines

•Naïve Bayes

•Linear Regression in both

ordinary least squares and

Gradient Descent approaches.

• Random Forests.

PROJECTS on Distributed and parallel computing.

Implemented a Scheduler in Pthreads c++ Library for High intensity Numerical Function:

•High intensity functions take a lot of time to compute so there is a severe necessity to reduce the computation time. One of the classical approaches to perform work decomposition is through multi-threading.

•Computed the Dependency Graph and look at some of critical Metrics in the systems like Width and critical path because they are essential to compute speed ups and no of processors to be used when performing parallel computing.

•Used techniques like Static scheduling, Dynamic scheduling with different chunk sizes and computed speed ups.

Implemented Massively Serializable Algorithms Like Prefix sum, Reduce, Merge Sort, Longest Common

Subsequence in parallel using OPENMP in c++.

•Implemented the classical Serial algorithms in c++ and computes how much time it is taking for different values of n from 10 to 1 billion elements and observed that there is a necessity to bring speed ups in the code.

•Implemented parallel Prefix sum, Parallel reduce, Parallel Merge sort, Parallel Longest Common Subsequence with a Strong Scaling experiment and observed dramatic rise in the speed ups with static scheduling and Dynamic scheduling for different chunk sizes. Implemented Search Engine for information retrieval course using TFIDF and page rank scores Used Hadoop Map-reduce in java.

•Implemented the Page rank Algorithm by google in distributed Fashion and performed 10 iterations to converge the algorithm on the whole Wikipedia data using Hadoop Map reduce.

•Implemented Term Frequency and Inverse Document Frequencies for Giga bytes of Data and sort all the words in Ascending order.

•Target is to retrieve the pages when a query is typed based on the high page rank and high tfidf scores.

•Used Machine learning algorithms like Logistic Regression combinedly with page ranks and perform classical tests on how fast the systems is responding to a query and Used L1-regularization for Faster querying.

Implemented Convolution operation on a very High

Dimensional image of 200 GB in a distributed Fashion in MPI (Message passing Interface in c++) and Matrix multiplication in CUDA.

•Implemented the Convolution operation in a distributed Fashion and handled all the communication on multiple computers in c++.

•Implemented the Classical Matrix- vector multiplication and matrix- matrix multiplication on Both MPI and CUDA to visualize the speed ups differences when using CPU and GPU.

Contact this candidate