Categorizing Webpages for Contextual Search (May 2011-October 2011): The Search-on-Select feature on the kikin app spawns a web search whenever the user clicks on a word or phrase from a webpage. Our goal was to filter the pages returned by search by category (business, arts, recreation, sports etc) to increase the relevance of the results returned. For example, a user might click the word "jaguar" from a page which is about automobiles, while another user might click the same word from a page about animals – kikin’s goal is to understand the user’s intent from the context - whether the user is interested in the jaguar car or the animal, and refine results accordingly. We are applying classification algorithms on the contents of webpages to categorize them, to make sure the categories of the search results filtered out by kikin match the category of the page the search was spawned from. In this example, the user who spawned search from the automobile page would be served with the prices and features of the most recent models of the car, while the user who spawned search from the animal page would be served information about the feline.

Roles played:

1. Retrieved the categories of 4,000,000 URLs from the listing provided by the Open Directory Project (ODP) to build the ground truth.
2. Created a sample of 40,000 URLs to perform initial tests. Initially, the categories with more URLs introduced a bias in the results, so we had to choose the sample carefully so that the expected number of samples from all the categories are same.
3. Extracted text from the URLs with Apache HttpClient, pre-processed the text to eliminate all stop words and function words, and developed custom libraries to convert each English word to the corresponding root word (e.g., "beautification" and "beautiful" are both stemmed to "beauty"). The library looks up a comprehensive English dictionary and uses Levenshtein distance as the distance metric while identifying the "close matches" of a given word.
4. Implemented methods for the creation and efficient storage of a high-dimensional sparse vector of tf-idfs of the various terms in the document pointed to by each URL.
5. Implemented the k-Nearest Neighbor algorithm to categorize the validation URLs first on Hadoop and later on commodity hardware using the Spring J2EE framework. So far, 60% of the validation URLs have been correctly classified, and the average time to classify each URL is less than 100 milliseconds.

Development of Web-based Dasdboard for User Behavior Monitoring on iPad (August 2011): The Search-on-Select app of kikin for iPad aggregates user behavior data on a per-day, per-user level and sends it to an application server hosted on AWS as a stream of JSON documents. The application enters the data to a PostgreSQL database, and the web-based dashboard, developed using the Google Charts API over the Spring framework, displays various graphs on user-base, usage, user acquisition and user retention for business analytics.

Roles played:

1. Designed the appropriate metrics for user behavior monitoring.
2. Developed and tested the application and deployed the same over AWS servers.

Extracting Demographic Information for kikin Users (April 2011-May 2011): The goal was to extract the gender and location information of the kikin users to offer them personalized deals while they browse for products in online shopping sites like Amazon or eBay.

Roles played:

1. Extracted the gender information from the Facebook accounts of the users, using the Graph API of Facebook. Implemented methods to process the results returned in JSON format.
2. Extracted the location information using the MaxMind API.