Jayanth Boddu

WHO estimates show that there are approximately 285 million visually impaired people worldwide, out of which 39 million are completely blind. It is extremely difficult for them to carry out many daily activities, one of which is reading - from reading a newspaper / magazine to an important text message from your bank. In 2010, Facebook launched a special feature that can help the blind use Facebook on their mobile phones. The feature reads out the contents of images that their friends posted on Facebook. A Demo can be found here. Converting an image to text can be viewed as two sub tasks - 1. Generating cpations describing the image 2. Text to Speech Conversion This project undertakes the task of generating captions to describe the contents of an image i.e Image Captioning. Image captioning is an important Image Processing task which is central to the task of Scene Understanding. A custom Deep Learning Model is built using Tensorflow and Keras based on Encoder - Decoder architecture with Bahdanau Attention Mechanism. The Model has been trained and tested on Flickr8k dataset. ImageNet, a CNN model pre-trained for image classification, is used for feature extraction. These features are passed through an RNN , which generates a caption , with due regard to previously generated words. This model improves the standard encoder-decoder architecture by the addition of an attention mechanism which helps the network focus on a part of the image ( instead of focusing on the entire image all the time ) based on the words previously generated. The generated caption is ranked based on Greedy Search and evaluated using the Bilingual Evaluation Understudy ( BLEU ) score. Neural Architecture in this project is based on Show, Attend & Tell,2015.

Built a 2D CNN + RNN model to accurately predict hand gestures to control a smart TV. A validation accuracy of 81% was achieved on a labelled dataset containing 30 frame videos. RNN+CNN and 3D CNNs were experimented with. The final model presented was 2D CNN (MobileNet architecture) + GRU. MobileNet architecture was chosen for low memory footprint (number of parameters) suited to webcam memory sizes. A complete set of experiments performed can be found here

Built a neural network from scratch and tested it on the MNIST dataset to predict digits from handwritten digits. Achieved a prediction accuracy of 87%. This was a hands-on exercise to understand the working of a basic neural network.

A case study on top 5 factors that contribute to churn among telecom customers using service usage based analysis of high value customers. This case study makes recommendations for a telecom company to reduce churn. In addition to data cleaning and pre-processing steps, this analysis contains a highly interpretable model built using Logistic regression and 3 high performance models built on a pipeline of decomposition using Principal Component Analysis and then modelled using Logistic Regression, Random Forest and XGBoost. SMOTE is used to mitigate high data imbalance and Sensitivity metrics are compared for all models.

Helped X Education select the most promising sales leads, i.e. the leads that are most likely to convert into paying customers. Built a model wherein we need to assign a lead score to each of the leads such that the customers with higher lead score have a higher conversion chance and the customers with lower lead score have a lower conversion chance. Potentially improved the conversion rate from 30% to ~80%. Reported the most important attributes that indicate high chances of conversion.

HELP, an NGO wanted to use their funding strategically so that they could aid five countries in dire need of help. The project used various clustering algorithms to group countries based on socio-economic and health factors to judge the overall development of countries to suggest 5 countries that need the aid the most. Also created interactive plots to explore countries in each cluster.

Business data of a hypothetical Bike sharing company was modelled to understand the effects of different factors on the daily demand. The most important factors that influence demand were reported along with an estimate of their measure of importance. Using this data, the business could predict demand with reasonable accuracy and use the insights to strategize business goals. This is an exercise in implementing Linear Regression models for prediction and inference.

Loan application data of a hypothetical bank was analysed along with credit history to report customer attributes, most indicative of default. Parameters related to loan default were determined and analysed using Exploratory data analysis techniques. This analysis is useful in estimating the risk of lending and thereby make decisions like loan rejection or hike in interest rate offered.

Completed a Coursera guided project on Exploratory Data Analysis of Breast Cancer Diagnosis data. This analysis helps in forming initial assumptions and a set of questions to give direction to further analysis. This is an exercise to identify variables potentially related to modelling breast cancer diagnosis, especially through visualizations of distribution of values and their variance.

IMDB was scraped for movie data. The data was analysed for insights like the most profitable movies,the worst performing movies and top 10 movie directors. This was an exercise in importing, manipulating and grouping data to answer relevant questions.

This analysis focuses on finding attributes that significantly affect wine quality classification and training a predictive model to classify wine quality into good and bad based on attributes. Analysis is pivoted on the variable 'quality'.

A simple interactive app to learn how tuning different hyper parameters influences the performance of predictions using Decision Trees.

A simple interactive app which looks at customer transactions of a hypothetical bank to derive insights on where the customer is making transactions.

Built an inventory tracking web app with all store related functionality using Full stack Javascript technologies. I used

• NodeJS backend
• ExpressJS web server
• Angular frontend
• MongoDB database

Built a web based biometric attendance system. Wrote C++ libraries to nodejs bindings to access device APIs. The device was connected to server hosting the web app to track attendance. Angular and bootstrap was used to build the web app. This app was used till the roll out of enterprise wide access control system.

For real time position tracking of high value assets, I developed a system involving low cost raspberry pi based scanners and BLE beacons. Beacons were attached to high value assets. Readers placed in each zone read realtime broadcasts of beacons and relayed data to a server connected through Ethernet. A web app showed real time location of high value inventory

This is virtual Internship offered by Quantium, an Australian analytics firm. Analyzed chip sales data for a hypothetical supermarket chain with 221 stores. Reported the most profitable customer segements and the size / brands of chips they prefer based on drivers of sales. Evaluated the effectiveness of store trial using metrics like increase in Customer turnout and sales. Presented all the results to the Category Manager.