Oct 2025 – Present · 7 mos · Bengaluru

• building models and systems for Real-time Ad Bidding

Jul 2024 – Jun 2025 · 11 mos

• Worked on Ad-personalization models for ShareChat and Moj. Mainly improving personalized Ads ranking models for SC & Moj, and maintaining features (batch and realtime).
• Worked on training distributed ranking models like MMoE, PLE and Two tower models for large scale embedding learning.
• Worked on Ad-Content based embeddings a lot to improve ranking models, and used LLM based feature generation to improve overall ROC-AUC like metrics, leading to overall ads value metric gains on platform.

Oct 2022 – Jul 2024 · 1 yr 9 mos

• Worked on improving Early stage Personalisation, boosting early stage successful watch rate from XX% to YY% and user retention by 0.X%. Tried improving content understanding of posts using CLIP and Multi-modal Large Language Models. Some key mentions:
• Improved Candidate Generation by doing improved cold-start of real-time trained FFM embeddings
• Created inference pipeline for CLIP-based embeddings at post creation, led it's adoption org-wide across several key teams.
• Trained several regression models to predict mature FFM embeddings using static-features for FFM cold-start. Improved p50 embedding convergence at post creation by 2X% abs
• Improved Topic Prediction using content and realtime features, increasing top-1 and top-3 acc abs. by Y% and Z% respectively.
• Automated AI-Generated Notification text for posts using Retrieval Augmented Generation based on Post content. Hosted Llama 70B model locally.

Oct 2021 – Sep 2022 · 11 mos

• Worked mainly in Trust and Safety team, Designed realtime moderation systems for streaming formats, tried to improve existing moderation models and Save moderation costs. Some notable mentions:
• Designed first version for real-time moderation systems for Moj Live, Sharechat Audio Chatrooms, and Improved Post moderation in ShareChat.
• Re-trained Visual moderation models using hierarchical classification to detect Pornographic, Gore, Spam & such categories. Improved overall recall@topK metrics, reducing Avg. views per unsafe post by XX%
• Created Toxic speech classification model using pre-trained speech embeddings; iterated to improve recall@top-K metrics
• Refactored Sharechat post moderation pipeline to move to more resilient job-based setup. Saved YY% of overall moderation compute costs.

Jun 2021 – Oct 2021 · 4 mos

• Working in the Multi-modal Content Understanding Team (Moj & Sharechat)
• Worked in developing better models for Content Moderation (NSFW, Bait, Spam) via Visual, Audio/Speech and Text modalities
• Worked on model inference, load testing, model serving, deployment/monitoring and continual retraining for models in production. Made efficient pipelines to handle inferences for multi-million posts per day.
• Optimized production inference pipelines to save ~90% of GPU inference costs for Content Understanding team. Received Spot Award for the same.
• Worked on the Google Cloud Platform stack.