In the Ads AI, I work on the problem of response prediction, in particular, predicting view-through rate of video-ads. We are investing heavily in building and leverage embeddings derived from the frames of the videos to help predict response on those videos. In particular, we are investigating using content embeddings derived from text, image/video of the ad to help in variety of prediction tasks in the Ads domain. To learn more about these applications please visit the KDD tutorial. More recently, my work has shifted to focus on building response prediction pipelines from scratch for the Stories product, where we find the most relevant ad to show in the Stories to drive revenue for LinkedIn.
Box Inc (Nov 2013 - Oct 2015), Software Engineer, ML Team At Box, I was involved in building the Machine Learning platform for improving user experience through data intelligence. My work involved analyzing usage data and content to come up with suitable ML algorithms for various product use cases. I was also involved in developing scalable and reliable software for these ML algorithms in Apache Spark. For a brief period of time, I was involved in supporting and building the data- analytics Infrastructure at Box, and thus, have accrued basic experience in technologies like Kafka, RedShift, Storm, Hive, Hadoop and ETL. Click here to learn more about my work at Box.
Google Inc (May 2010- Aug 2010), Software Engineering Intern Streetview panaromic views are heavily dependent on the pose of the vehicle as captured by the GPS. In big cities, the GPS data is grossly erroneous. In such situations, images can be used to correct the pose of the vehicle offline. The pose can be corrected by using a computer vision algorithm called loopclosing. The candidate intersections are pre-computed using GPS information and transform compute to align the vehicle pose at these intersections. The alignment is used to perform optimization in order to estimate the true location of the vehicle. My internship at Google was to develop and test appearance-based loop-closing algorithms on street view data, and suggest evaluation measures of performance of the algorithms.
Nvidia (Aug 2006- June 2007), Software Engineering After the successful completion of MS in IIT Madras, I joined the video decode display team at Nvidia Graphics India. The work involved writing and modifying driver code at the kernel level in order to implement functionalities like post processing, image scaling, in-loop de-blocking using the GPU (Graphics processor unit). The job required one to learn or be aware of the various video-coding standards like MPEG-2, MPEG-4, WMV, H.264 and VC, knowledge of Windows OS like Windows XP, Windows Vista, co-processor architecture, low-level C programming