Ravindra Kumar

Data Scientist @Times Internet

Noida, IN

Recommender System• Built personalized feed and push notification recommendation systems serving 1.7M+ users across iOS and Android.• Designed candidate generation using WALS matrix factorization and ScaNN-based ANN search over 50K+ articles.• Developed ranking and re-ranking models with click normalization and position bias correction to optimize CTR.• Added a sampling-based exploration strategy in production to preserve content diversity and mitigate selection bias.• Productionized ML pipelines using TensorFlow (candidate generation) and PyTorch (ranking); orchestrated via Airflow.• Delivered 120% CTR uplift with stable DAU/MAU and +15% scroll depth at < 200 ms P95 latency on FastAPI/Kubernetes.2. Editorial Judgement System• Designed an LLM-powered editorial decision layer for feed and push distribution, aligned with human editorial standards.• Extracted structured editorial signals using LLaMA-3.1-70B with SFT and QLoRA for semantic content understanding.• Integrated LLM features with ML ranking and rules, achieving additional ∼ 80% CTR uplift and ASPIRE Award recognition.3. Agentic Editorial Knowledge Retrieval System (Archive RAG)• Built agentic RAG over TOI (4M articles) using Milvus + OpenSearch hybrid retrieval with cross-encoder reranking.• Designed metadata-aware retrieval with section and temporal routing for background, latest, and timeline editorial research.• Enabled LLM inference via vLLM (LLaMA-3.1-70B) with citation-grounded generation to ensure archive-first responses.• Deployed LangGraph-based agent pipeline in production on Kubernetes, reducing manual archive lookup for editors.

Chaudhari Nihal Singh Inter College, Aimi Pilibhit (UP)

Intermediate Education

Indian Institute of Technology, Kharagpur

Bachelor of Technology, Ocean Engineering and Naval Architecture