Apr 2023 – Present · 2 yrs 11 mos

• Worked on User / Item personalization problem, specifically the Candidate generation layer for serving millions of users.
• Feed AI:
• Recall models: Matrix factorization, Content Embedding based Candidate Generator (CG), User state model-based CG (partial contribution)
• Topic-coherent CG: Based on an interacted topic by the user, suggest similar videos to capture the user interest. Similar posts were recommended by learned topic embedding for a given post.
• Recent-aware CG: In-session personalization kind of CG, where it would account for user's last k interaction and based on that similar items would be suggested. Similarity is measured by a real-time learned vector representation of user and item based on their interaction.
• Two-tower CG: Incorporate different user and post features into CG to capture users personalization better on short-term as well as long-term aspects.
• Cost Optimization::
• ~20M+ INR / quarter savings were observed by doing various optimizations at job/service - deployment level in Content Understanding pipelines.
• [Reasons]: By reducing the data transfer cost from one region to another, setting up proper HPA metric, CDN cache fill / CDN lookup cost savings, Using GCS bucket client API instead of using public storage url, using spot compute instance instead of reserved instance

Aug 2021 – Apr 2023 · 1 yr 8 mos

• Working on ML challenges around Content and User representation, Trust & Safty of platform: Moderation AI and Feed AI: personalization and recommendation for Moj, ShareChat & MX Takatak (350M+ Monthly active users).
• Have worked on several projects, here are few of the key contributions:
• Content and User representation:
• Face clustering algorithm
• Gender/Age/count-of-people detection
• Continual/Online learning for Topic prediction
• Indic Language Identification
• These representation can ultimately be useful in Feed AI for personalized recommendation and user retention.
• Moderation AI:
• NSFW detection models/Improvements in the models for ShareChat, Moj, Livestream (currently 1M+ daily active users)
• NSFW video Explainability model for finer annotation
• Improved model has an Impact of cost saving of >4M+ INR per quarter in terms of Moderation cost.
• Analyzing content came on the platform Sharechat and Moj as a part of the Content and User Understanding team, wherewith the help of ML/DL and statistics will try to get insides of content/user which can eventually help in various User retention aspects.