Sample Projects

Classifying Retinal OCT Images

Can we detect Retinal Disorders in OCT images using CNNs?

Card image cap
Card title

Challenges: Occasional poor image quality, differing image heights and widths, unbalanced classes

Methods Applied: Filters to reduce noise, fixed image pixel sizes, balancing classes, and training with blocks of 1000 images at once

Utilities: Python, Keras, Tensorflow 2.0, OpenCV, CNNs, Pretrained and custom models

Result:  Achieved a 97% testing accuracy on average with 5-fold cross validation with pretrained models on a class-balanced dataset. Training time reduced by 80% compared to a traditional non-pretrained CNN modeling approach.

 

 

Recommending Movies

Given a user's movie preference profile, can we recommend movies?

Card image cap
Card title

Challenges: Distributed data, Messy data, Imbalanced categorical variables

Methods Applied: Database joins; Data cleaning and splitting into clean variables; One Hot encoding

Utilities: Python, Pandas, Seaborn, Content-based filtering, Collaborative-filtering

Result: Able to generate the top 'X' movies suggestions based on user profile and ratings, least 'X' recommended movies, Most popular 'X' movies lists

 

 

Predicting Loan Default Status

Can we predict who is most likely to default on their loans using demographics and payment history?

Card image cap
Card title

Challenges: Unbalanced classes, Messy Data

Methods Applied: Data wrangling, balancing classes for training and validation

Utilities: Python, Pylab, Scipy, sklearn, Jaccard_similarity_scores, KNN, Decision Trees, SVMs, Logistic Regressions

Result: Achieved a preliminary prediction accuracy of 75% in determining loan defaulters 

 

Three-way Tweet Classifications

A tutorial for three-way sentiment analysis using standard python NLP libraries such as NLTK.
Card image cap
Card title

Challenges: Unlabeled data, raw tweets with handles and unwanted data for sentiment analysis, misspellings, lemmas and stem words

Methods Applied: Tokenization, stemming, lemmatization, vectorization, PCA, TSNE, Naive Bayes Classification

Utilities: NLTK, Naive Bayes Classifier

Result: TSNE method validated with labeled data with an accuracy of 76%. The same method is used to derive sentiments on unlabeled raw tweets.

 

League of Legends Tweet Sentiments in real-time with Dash.

Can we track Tweets associated with tags/keywords to do real-time sentiment analysis?

Card image cap
Card title

Challenges: Problem specific data collection, continuous data-stream from tweets, user perception of product or people or brand

Methods Applied: Data collection from Twitter feed with keyword matching for the brand or products, real-time tweet cleaning, storing into an SQL database, real-time sentiment analysis, translation of non-english tweets to english, live interactive dashboard template with "Dash" to create a web applet to stream the data and the sentiments associated with each tweet         

Utilities: Python, Tweepy, Textblob, sqlite3, json, twitterstream, NLTK, SentimentAnalysis, Polarity scoring, Subjectivity

Result: Demonstrated an ability to gather, parse, clean, and understand the real-time sentiments associated with League of Legends new champion launches 

 

 

Tracking and bot-investing with Robinhood Stocks.

Can we track, analyze, and auto-invest on Robinhood?

 

Card image cap
Card title

Challenges: Accessing real-time stock information, automating purchase decisions to maximize profit

Methods Applied: Webcrawler to gather near real-time stock info from Yahoo.com, Financial analysis to generate candlestick visualizations, short-term, and long-term and golden-cross indicators for predicting profitability of each stock

Utilities: Python, Robin_stacks, numpy, ta, tradingstats, dash, plotly, pandas

Result: Automated the process to track, analyze, and automate Robinhood stock portfolio management with python that can make guided purchase decisions to reduce time investment and improve purchase decisions for deriving profits. 

 

 

>



4

Publications (1 published, 1 in press, 2 in prep)

12+

Data Science projects and counting (check my github repo)

8+

Years of Work Experiences spanning healthcare, human behavior, and data science

About me

I use AI and data driven decisions to personalize UX, Learning, and Cognition. I treat data science as a form of art that meets science. I am excited about deriving actionable insights from data to improve your products, your customers' experiences, and the world.

Download CV

My Work Experiences

I work across multiple domains and integrate them seamlessly to better understand human behavior

Data Science

I work on a variety of data science projects that utilize machine learning, AI, and statistical modeling applied to learning and behavioral predictions, user segmentation, computer vision in healthcare, risk analysis, and sentiment analysis

Back to top

UX Research

I conduct research and create end-to-end research solutions to understand learning and human behavior in a variety of contexts. I am motivated by the work geared towards optimization, inclusion, and personalization of user experiences across domains -- technology, entertainment, healthcare, and learning.

Back to top

Visioncare Experiences

I have 3+ years of clinical experience as an Optometrist in India. My current research and future endeavors are tied to my experiences as a clinician. I believe that every individual should recieve a personalized experience when interacting with technology, entertainment, and everyday utilities. I want to understand the needs of the next billion and promote equity, inclusion, and universal design through my work.

Back to top
BITS, Pilani
L V Prasad Eye Institute
U C Irvine
U C Riverside
U of Michigan