Palo Alto, California, United States
SWE @ Google Maps - Building a 3d model of the real word.
Working in Google Maps to create virtual 3D world.
- Built an end-to-end ML pipeline to rank merchant posts being served on Google Maps as a response to COVID-19 to help businesses share their updates with their customers. - Designed and sent Side-by-Side surveys to human raters to generate labels for point-wise and pairwise models. - Analyzed the survey results and applied NLP to decide the overall sentiment and quality of the posts. - Built a MapReduce feature extraction pipeline that reads from multiple sources to produce TFExamples for training. - Used TFRanking to set up a Group-wise Scoring Function(GSF), and a DNN to train and rank merchant posts. - Improved the accuracy of the ranking model by 11% and launched the feature in production.
Developed a feature computer that serves trip type scores per hotel to improve the TensorFlow DNN model for Hotel Search ranking. Designed and implemented a datastore for Transit related features: - Wrote a parser to filter the shortest walking duration & frequency of each transit type (bus, train, etc.) within a 10-minute radius. - Transformed raw transit data into semantically meaningful features and stored them as protocol buffers. - Developed a feature computer to process protocol buffers and improve hotel ranking in transit heavy locations (Tokyo / New York). - Improved the server load time by reducing the RAM cost from 700 MB to 16 MB. Designed an algorithm to score hotels based on its proximity to transits and implemented a twiddler to adjust the ranking results. Implemented a program that backfills each TensorFlow example with experimental features and eliminates training/serving skew. Added features to distinguish business category hotel searches from entity searches.
- Implemented a virtual interplanetary Delay-Tolerant Network (DTN) using NASA's Interplanetary Overlay Network (ION) package. - Used this network to develop a cached website that supports DTN. - Wrote a transfer module in bash that communicates cached signals between the client and the webserver to handle non-real-time requests.
- Creating a new Alt-Coin by forking Bitcoin's source code. - Use this cryptocurrency to create a genesis block and start a new blockchain. - Building on top of this blockchain to expand the usage of blockchain to other logistical transactions.
-Developed a web application which serves as an internal debugging tool within Google ads infrastructure: - Provides the ability to dive into any entity's extraction timeline as it is fetched from the database, indexed into logical shards and written to multiple output formats. - Tracks statistics related to processing time and delays during scheduling & merging of any entity under full extraction. - Inspects real-time info changes on batches and workers for an entity under incremental extraction. -Wrote a modular full extraction library in C++ which automated a previously manual process, reducing the average turnaround time from ~1-2 hours to ~10 minutes. -Added feature to filter shards and task execution info for any entity by reading corresponding bigtables and shard metadata. -Added another feature to read intermediate snapshots to retrieve any entity record in channel shards for the provided protocol buffer and field. -Built and hosted a C++ backend server to handle HTTP request and responses for both real-time and periodic extractions. -Co-authored the design document, specifically by designing the backend server and workflow of each HTTP request-response and implementation for full extraction libraries. -Attended weekly ML Talk Series & Leadership Development workshop. -Tech used: C++, Server Configuration, Web development, Bigtables, Protocol Buffers, JSON