Infrrd.AI

Software Developer Intern• May 2018 - August 2018

•    Trained a Recurrent Neural Network and Convolutional Neural Network based encoder-decoder sequence to sequence model with Bidirectional LSTM and attention using Keras to perform spelling correction as a post-processing stack of tesseract-based table extraction from images/pdf. Trained the model using Google Cloud with GPU and achieved an accuracy of 98% on text output from OCR table extractor.
•    Implemented a tree-based fuzzy match clustering for OCR documents to group correct and incorrect words and extract domain vocabulary. I came up with a unique noise generator to generator noisy data-set, which mimicked the real-world errors.
•    Built a POC from scratch for classifying image/pdf document type and extract data with 95% accuracy using Google Vision, word-based pre-processing and neural network for 5 document types. Deployed it as a RESTful API using python Flask on AWS.

Citrix India R&D Pvt Ltd.

Software DevTest Engineer 1• July 2016 - July 2017

•    Interacted with Platform and Networking Dev-Test team to solve Load-balancing problems and developed 75% Automation scripts for Cluster, High Availability, VLAN, SDX, LowerPE, PPTP, Telco features of NetScaler VPX and its Cloud deployment.
•    Led a team of 8 members by being Scrum Master and managed to conduct daily sprints with 6 test case automation rates per engineer per day

Citrix India R&D Pvt Ltd.

Software DevTest Engineer Intern• January 2016 - June 2016

•    Proposed and implemented an efficient way to use less resources and achieved 80% automation of RIP with IPV6 addresses in NetScaler

Secure, Offline Feedback to Convey Instructor Intent

2015 IEEE Seventh International Conference(T4E)• Warangal, India

• Presented a IEEE paper named "Secure, Offline Feedback to Convey Instructor Intent" at Technology for Education(T4E) along with a professor and 3 other authors, 2015 IEEE Seventh International Conference held in December 2015 at Warangal, India
• Unique way of storing solutions in memory. Students answers are minimized to minimized DFA and again BFS is applied and hashed. Hashed are compared to check for correctness of student’s answer

Abnormal Event Detection

PES Institute of Technology• January 2016 - June 2016

• Learning algorithm converts normal event patterns video input into time-varying 3-dimension frames used to learn a set of features.
• Learnt features are used to reconstruct test videos and the reconstruction error is analyzed to detect abnormal patterns.

Moderator Detection in a debate by Speech Processing

PES Institute of Technology• January 2015 - June 2015

• Trained various Hidden Markov Models on MFCC audio vector of people participating in a debate and moderator speaking style.
• Test audio clips were analyzed at a rate of 5sec and summarized to predict the portion of audio and the person who moderated the debate.

Adversarial Search based Game Playing Agent

University of Southern California• August 2017 - December 2017

• Implemented min-max agent to play a candy-crush like game using artificial intelligence.
• Optimized the run-time efficiency by using an alpha-beta pruning version of the algorithm.

JFLAP Quizzes with secure offline self-assessment

PES Institute of Technology• January 2015 - June 2015

• Unique way of storing solutions in memory. Students answers are minimized to minimized DFA and again BFS is applied and hashed. Hashed are compared to check for correctness of student’s answer
• Presented a IEEE paper named "Secure, Offline Feedback to Convey Instructor Intent" at Technology for Education(T4E) along with a professor and 3 other authors, 2015 IEEE Seventh International Conference held in December 2015 at Warangal, India

DDOS Attack Detection and Prevention in Pluribus Switch using SDN

PES Institute of Technology• January 2015 - June 2015

• Source and destination IP of each packet is analyzed to predict a possible DDOS attack
• VFLOW feature is used to block the source IP thus preventing any further DDOS attack

Optical Character Recognition using Neural Nets

PES Institute of Technology• March 2015 - June 2015

• Neural Network model is used to recognize individual character in an image using labeled character images

Smart Home Automation

PES Institute of Technology• January 2015 - June 2015

• Web app controls lights by providing an interface to control voltage between smartphone and raspberry pie