Sep 2023 – Nov 2024 · 1 yr 2 mos

• Realtime Ranking and Personalization
• · Working on building real-time ranking capabilities for user-feed personalization on Moj and Ads.
• · Responsible for cost-reduction of 30% by moving to spot instances for training workflows.
• · Reduce training GPU cost by implementing mixed float operations and loss scaling.

Oct 2021 – Sep 2023 · 1 yr 11 mos

• Moj Feed AI, Recommender Systems, Notification Personalisation and Volume Optimisation
• · Building infrastructure for personalised notifications capable of handling scale of millions of users.
• · Ran A/B experiments on personalised notifications as well as volume optimisation to achieve CTR gains and re-activation of users.
• · Achieved 80% savings in compute and SQL costs through various optimisations on infrastructure and storage.
• · Worked on solving Post Cold start problem for Moj.
• · Achieved a 0.X% retention gain in D1 users.
• · Increased the number of posts served to users by 4x, leading to improved diversity and post exploration.
• · Built early stage candidate generators targeting early post discovery to maximise chances of success.
• · Improved personalisation of posts in early stage to push for quick post embedding convergence.
• · Bridging the supply-demand gap in smaller Indian languages like Gujarati, Bhojpuri, Rajasthani, etc. through specific candidate genarators catering to these business usecases.
• · Running and analysing various A/B Experimentations of user and post variants to test different feed recommendation logic to optimise for metrics like user retention, video play time and engagement metrics.
• · Responsible for multiple core feed microservices and training pipelines and enabling autoscaling as well as assuring complete uptime of our feed models.
• Mentors - Hastagiri Vanchinathan, Director(Feed AI/Ranking), Michal Porvaznik, Staff ML Engineer and Jan Ruziev, Staff ML Engineer.

Apr 2021 – Sep 2021 · 5 mos

• Recommender Systems, Feed AI, Duplicate Detection, Google Cloud TPUs
• · Worked in the modelling team of Moj Feed AI.
• · Worked on online recommender system leading to significant gains in business metrics like retention, video playtime, feed quality etc.
• · Built scalable, robust core and cost efficient modelling pipelines for our Feed AI Ranking component.
• · Worked on model inference, serving, load testing and deployment/monitoring for our core feed models.
• · Worked on Large scale nearest neighbour search engines for Feed Ranking and Content Understanding.
• · Extensively involved in developing and running multiple A/B Test Experiments on User and Post Level business metrics like retention, playtime, engagements etc.
• · Responsible for various end-to-end machine learning pipelines for Sharechat and Moj.
• Mentors - Hastagiri Vanchinathan, Director(Feed AI/Ranking) and Peddakota Vikash, Lead Data Scientist

Aug 2020 – Mar 2021 · 7 mos

• Factorization Machines, Recommender Systems, Personalisation, Feed AI, Duplicate Detection
• · Worked on Moj’s Recommender System using FFM from scratch.
• · Worked on hypertuning training parameters for Recommender Systems to optimise feed quality.
• · Worked on online recommender system leading to significant gains in business metrics like retention, video playtime, feed quality etc.
• · Scaled Nearest Neighbour Search to 100 million vectors for realtime duplicate detection.
• · Developed Fast 3D Video models to generate rich video representations.
• · Improved on existing Topic Hierarchy Classification and Tag Prediction models for better content understanding.
• · Built tools to visualise Feed Recommendation.
• · Responsible for various end-to-end machine learning pipelines for Sharechat and Moj.
• Mentors - Debdoot Mukherjee, Vice President(AI)

Jan 2020 – Aug 2020 · 7 mos

• Multi Modal Learning, Feature Extraction, Unsupervised Learning, Large Scale Similarity Search
• · Built a multi modal learning system(text, audio, video, images) from scratch that is used for Tag Prediction and Tag Filtering over a spectrum of 15+ Indian languages.
• · Worked on efficient and fast feature extraction for video and audio content on the Sharechat content platform to act as a base for multi-modal learning and graph learning.
• · Worked on duplicate detection of posts on the content platform and reduced required resources from 48 EC2 machines to just 1 machine by optimizing the similarity search and feature space by a factor of 16x.
• Mentors - Debdoot Mukherjee, Vice President, AI and Doney Alex, Lead Data Scientist

May 2019 – Jul 2019 · 2 mos · Gurgaon, Haryana, India

• Decoupling Localization and Classification in Single Shot Temporal Action Detection
• Working on developing a robust model performs action localization and classification using a single end-to-end trainable procedure.
• Mentors - Ankit Agarwal, CEO and Ravikant Bhargav, Research Head

May 2018 – Jul 2018 · 2 mos · Work from Home

• Creating a chatbot using Google DialogFlow API to bridge the gap between buyers and sellers in the shipping industry.
• Developing a model that Dynamically prices Hotels for Maximum Revenue.
• Worked on developing the idea of Supply Chain Management.
• Mentors - Gaurav Sharma, CEO

May 2018 – Jul 2018 · 2 mos · Bangalore

• Worked on a project on Rapid 3D Reconstruction of UAV video stream for Centre for Artificial Intelligence and Robotics(CAIR), a lab of DRDO.
• Mentor - Dr. Malay Nema, Scientist at DRDO