San Francisco Bay Area
ML at Google Maps, Visual Lanes. Working on Roads Data Quality, improving bus lanes detection and lane connectivity on all road priorities by using SoTA Computer Vision modeling on Satellite, Aerial and StreetView imagery.
ML at Google Search, Content Understanding - Lead quality efforts to improve scoring/relevance model for named entity recognition system for Google Search which annotates and scores knowledge graph entities on XXB indexed documents. - Fine-tuned a SOTA Gemini Encoder model which acts as a teacher model for a BERT based student model. Performed successful teacher distillation to student, preserving most of the teacher model improvements. - Led efforts to reduce serving latency for Lens query generation model. Experimented with various serving optimization to reduce serving latency by ~50% on XXXM traffic.
Part of Knowledge Engine Quality team in Google Search, improved search ranking and relevance. Projects undertaken - Metrics correlation and forecasting: Did an impact analysis of Document Understanding system of Knowledge Graph on Search. Established a correlation between Document Understanding metrics and Search metrics to drive future launches to get an estimate of improvement in Search. - Short Video Ranking: Improved ranking of Short Videos being displayed on Search Result Page by using a similarity match between annotated/ contextual KG entities for a query and KG entities for retrieved documents. - Embedding based Document Understanding and Nearest Neighbor Vector Search: Created a system for annotating KG entities in a document which used a Dual Encoder architecture to create document and KG entities embeddings. The document embeddings were used to perform a nearest neighbor search on an index of all KG entities to retrieve the closest entities relevant to the document.
• Worked on designing and implementing API and services for KYC process of sellers, assisting compliance team to verify documents submitted by sellers, thereby increasing velocity for seller registration. • Did impact analysis for legacy APIs for the initiative of sellers registering for various marketplaces. • Actively involved in On-Call tasks for bug fixing and unblocking registering sellers.
• Developed data ingestion pipeline for external data to be processed and consumed by internal research orgs. • Contributed in development of a uniform data ingestion framework used org wide. Built an efficient config parser along with test driven development. • Developed an efficient tool for scraping and parsing open source data and making it available for downstream users
•Documented optimal client configurations for various databases in a polyglot cloud environment. •Parallelized CI/CD builds thereby reducing the build time by 3 times. •Developed and integrated the Cassandra endpoint in the framework. Built the ORM for Cassandra and wrote robust unit tests.
• Automated ETL generation, table creation and table alteration by templatizing the respective YAML files. It resulted in reduction of number of lines of code by 10 times, reduction in touch points for code modification, easy debugging since the number of lines have been reduced significantly, could be used to validate legacy ETLs, less engineer hours being spent to modify the existing ETL • Built an ElasticSearch backend for the Global Data Organization. Designed and wrote the schema for the ElasticSearch indices. The indices were to be populated by parsing the respective Confluence pages. Created a queuing mechanism, using minimum API calls, to parse subsequent child pages when given a parent document, created an efficient insertion and updation mechanism of data into the ElasticSearch indices. • Designed an algorithm for efficiently validating the difference in data between two databases. Came up with a hashing based join operation to validate the data present in both the databases and coded it in PySpark. Wrote a multithreaded program in python for simultaneously running query on both the databases.
• Worked as a Research developer, developing and deploying neural network models for different customer projects. • Worked on Protein Folding problem, understanding the problem and it’s complexities. Implemented a differential neuralnetwork in PyTorch and performed various ablation studies. • Worked on developing a Child Pornography Detection Pipeline for North Rhine - Westphalia government. The pipeline was a two staged process,first developed a model for pornography detection implemented in Fastai on PyTorch backend then implemented a pre-trained model for multiple face detection and age estimation. The out put from two models were combined to make a final prediction. • The prediction pipeline was then deployed as a Docker Container and shipped to the client.