Siva Sankara Rao Kongara

Implemented a deep learning network to train a Lunar Lander in OpenAI Gym environment using Tensorflow.
Built a fully connected single hidden layer network to train the network and I have used replay memory to randomly sample experiences for the training.
Saved the trained model and evaluated the bot with random seeds and achieved an average score of 180.

In this project, I have implemented a program to identify the speaker's gender.

The speech was recorded and a Fourier transform of the analog signal with respect to time was taken. Then, the primary frequencies in the signal were found in the Fourier transform and based on that the pitch was identified.
MATLAB was used to code the entire project.

Women tend to have high pitch and thus high frequencies and vice-versa for men.

Implemented a two hidden layer neural network for the task of classifying handwritten digits.
Implementation was in Python without using any advanced libraries.

Incorporated Dropout regularization scheme to improve the accuracy of the network to 99%.
Softmax activation was used at the Output layer and Cross Entropy loss was backpropagated.

Analyzed the New York city taxi dataset to locate pick-up/drop-off hotspots using Apache Geo Spark.
Deployed a three-node Hadoop and Spark cluster to analyze the dataset.
Performed spatial queries using Spark SQL on the cluster to identify the Z score of a cell and thus its rating compared to its neighbors.
Carried out performance analysis on various sizes of datasets and node configuration using Ganglia, a network monitoring tool.

Built a visualization tool to effectively analyze the spatiotemporal trends in crime across the city of Chicago.
Used PostgreSQL to pre-process 6 million records and prepare the data for the visualizations.
D3 library was used to visualize various spatial and temporal data and jQuery was used to render other parts of the visualizations to aid with filtering.

Managed a team of seven to implement a graph Database Management System (DBMS) from a relational DBMS.

Created Node Heapfiles and Edge Heapfiles to store the node and edge information.
Modified the comparators and the iterators to work with the new Node and Edge data types.
Programmed the database system to efficiently perform batch insertion/deletion of Nodes and Edges into the database. Created a new Z-tree index to efficiently index higher dimensional data in the… Show more

Managed a team of seven to implement a graph Database Management System (DBMS) from a relational DBMS.

Created Node Heapfiles and Edge Heapfiles to store the node and edge information.
Modified the comparators and the iterators to work with the new Node and Edge data types.
Programmed the database system to efficiently perform batch insertion/deletion of Nodes and Edges into the database. Created a new Z-tree index to efficiently index higher dimensional data in the nodes.

Processed queries to fetch nodes and edges with and without using an index with various Order By clauses.
Programmed complex join operators between Nodes and Edges to traverse through the graph using the new join queries.
Java was the programming language used.

Show less

Implemented and deployed a position mapping application using iBeacon technology.
This application was used to map a user's location on a given map based on the signals received from various iBeacons located.
Made used of the strength of the signal and the location of the iBeacon to approximate the person's position.

The goal of this project was to reduce the network congestion in random access networks.

I have analyzed a bio-inspired learning technique to reduce packet collision in multi-node random access networks. I have used a probabilistic model for learning each node's probability to send a packet based on its past experiences and thus incorporating the network's knowledge before transmission.

Various link layer back-off techniques have existed to combat the issue of network memory… Show more

The goal of this project was to reduce the network congestion in random access networks.

I have analyzed a bio-inspired learning technique to reduce packet collision in multi-node random access networks. I have used a probabilistic model for learning each node's probability to send a packet based on its past experiences and thus incorporating the network's knowledge before transmission.

Various link layer back-off techniques have existed to combat the issue of network memory issues and line fading. I have investigated whether combining these techniques with the bio-inspired learning could increase the throughput. I have incorporated geometric back-off technique to simulate this idea and have achieved a better throughput for a two node network.
Show less