We have developed and patented a cooking robot that can cook all kinds of Indian foods in a completely automated manner.Our prototype became the first machine to automatically cook Indian dishes (it was demonstrated in public for the first time on November 17th, 2017).We built the prototype with a grant from the Department of Science and Technology (through IIT-Bombay) in 2017.

Our research lab Aiaioo Labs focuses on natural language processing and machine learning. For the last year and a half, we have worked on deep learning algorithms and developed a curriculum for teaching deep learning through a series of coding exercises: https://github.com/aiaioo/DeepLearningBasicsTutorial/ In terms of research, we developed a new kind of classifier called the Perplexed Bayes classifier that is an improvement on the naive Bayes classifier, were one of the first research groups to investigate intention analysis, a technology that augments sentiment analysis, and have investigated deep learning models.We've also been contributing to open source. We donated a machine learning algorithm (a rock-solid implementation of the Naive Bayes classifier) to Apache OpenNLP (wrote the OpenNLP code for supporting this type of classifier). We also developed a framework for Arduino programming in Kannada, Hindi and Tamil.We've also published research papers on programming without syntax in English, Tamil, Kannada, Hindi, Japanese, French and German.

I worked on natural language processing technologies for under-resourced languages, resulting in a publication in EACL 2009: 'Large Coverage Root Lexicon Extraction for Hindi'.

I worked on an advertising platform and its analytics consoles. I started working on Ruby on Rails and was very fortunate to be allowed to take Dan Klein's Natural Language Processing course at UC Berkeley.

MobiTV developed apps that allowed television programs to be streamed to mobile phones.I worked on integrating billing systems for various mobile phone carriers to enable premium in-app purchases.I built enhanced program guides to be downloaded into MobiTV apps that had already shipped on mobile phones.I put together a skin fitment checking tool which was a simple XSLT transformation applied to skin offset parameters supplied in XML format. The XSLT would render the XML into a webpage in a browser.I remember camping out with colleagues on Mount Shasta on a weekend testing TV reception on a mobile phone 11,000 feet above sea level.I also got to take M. I. Jordan's 'Statistical Learning Theory' course.

I spent my last year with IBM in a travelling job as a Master Data Management Consultant for Software Group. I served as a technical consultant on the Trigo Product Server (later renamed the Websphere Product Server), a solution for Product Information Management. In that role, I was asked to help transition a large retailer, and was able to speed up some migration steps from a few days to a few hours.It was also my first time working on the West Coast. I moved to California to take up the job on a 250 cc motorcycle, packing my bags and a tent on the back seat and riding from Raleigh to Burlingame over 18 days (covering a distance of 3600 miles).

I developed mobile device UI libraries (AWT for Personal Profile and so on) for the J9 VM. I worked on both Nokia and Motorola devices and on a phone I don't remember the name of that ran Linux and used Qt for its UIs. I also wrote an image format decoder (GIF/JPEG/PNG) for use in the mobile phones. I also recall having developed a tool to debug memory leaks in embedded hardware. (Many years later, in Bangalore, India, I met an IBMer carrying a phone that our team had adapted a J9 VM for).I also applied for a couple of patents (for IBM) for a technology for turning stand-alone Java programs (say one written using AWT or SWT) into web applications (or collaborative web applications) that could be run on Tomcat (and they would need no modifications whatsoever). The technology was an alternative to GWT.My last patent at IBM was for a collaboration server that could take a stand-alone Java program - like a two-player chess game written in AWT - and turn it into a collaborative web application semi-automatically (and sometimes fully automatically) so that two users could access it on different computers and interact through it over the web. And this would be a web app that could be run on a generic app server like Tomcat (with some JVM-level tweaks).

I worked as a tester on a product called CMVC (a CVS-like code repository used internally at IBM) among others. I applied for a patent for a test tool that could estimate the amount of time defects had gone undetected in code.I built a fun little Wiki plugin for the team (on TWiki) that had a parrot that would pop out of the page and read out all the manual test-cases one by one so the testers could follow the instructions without having to read the steps while executing them. I learnt a lot about large organizations at IBM.I also created a simple Eclipse plugin for a concurrency testing tool developed by researchers at Haifa labs.And finally, I tried taking the concept of eXtreme Programming and applying it to testing. I explored - with the encouragement of a brave testing team - a process called eXtreme Testing that was inspired (as Agile Processes are) by Kanban and Kaizen (Japanese just-in-time manufacturing).The idea was simple. Since defects cost more to fix the longer they are in code, one can in theory lower the cost of rework and redesign of software, reduce risks, and keep projects on time if we test software as early as possible (in some cases even before the features are complete). TDD is the ultimate in early testing, but it requires developers to do the unit testing. What we came up with was a compromise for firms with separate development and test teams.

I was invited to join IBM by folks at the Smalltalk group of OTI Labs (a startup that IBM had acquired) and with whom I had had the pleasure of working in 2001 (at the 'Country Club' - the little OTI office space complete with barbecue grill and pool table behind the NC State Vet School).I developed a 'JSP' engine for VisualAge Smalltalk, a mechanism to integrate sessions across two incompatible VA Smalltalk libraries for server-side programming (WebConn and Server Smalltalk), and developed failover mechanisms using Websphere Application Server's IP Sprayer. I also integrated WAS's session persistence mechanism with VA Smalltalk's session management system so that Smalltalk servers could be operated as horizontally scaling clusters inside WAS.

I worked for the IBM VA Smalltalk group (OTI) on a JSP engine for Smalltalk and on load balancing mechanisms integrated with WAS.Later, I integrated the WebConn and SST frameworks, in order to leverage the failover enhanced session management subsystem in the SST framework.Still later, when IBM decided to make Java the language for middleware, I created a proof-of-concept language bridge (using JNI and Smalltalk's native functions) that allowed Smalltalk programs to use Java libraries.

I was trained to work on the Sun Internet Mail Server and Netscape Application Server (iPlanet). I recall having learnt a lot about the IMAP, SMTP and LDAP protocols, and about how to make software scalable and reliable.

Developer and tester on this UK-based website: http://www.widelearning.com. I developed the installer for the software, did a little bit of Java J2EE coding, and helped test the product.

I developed a control system, protocol (language), noise filtering algorithms, and resolution enhancement algorithms for a remotely operated Gamma Ray Burst observation telescope.The three novel features in that control system were:a) sensor resolution enhancement - the use of statistical methods to enhance the resolution of a low-resolution sensor (we increased the resolution using the central value theorem by averaging sensor readings affected by random gaussian noise).b) noise elimination algorithms - we had to classify each new data point from the sensor as being either a valid reading (to be used in estimating the real star position) or a noisy reading (to be discarded) and eliminate as much of the noise as possible.c) generalized PID (proportional, integral, differential) controller equations in altazimuth coordinates - we computed and used generalized digital control equations in spherical coordinates that allowed an altazimuth mounted telescope designed for any latitude to track a star from an observatory located at a different latitude from the one the telescope was designed for.We used to develop and test the software on a simulator in the IIA office in Bangalore, then drive up the hills with the liquid nitrogen canister transports and spend nights evaluating the tracking system on a real telescope frame.Real-life hardware can be unforgiving. On one occasion, a missing minus sign caused a giant telescope to turn on its head and crush an expensive piece of recording equipment. It was entirely my fault.