Hire Surya K. - Fine-Tuning Engineer

* Developed an end-to-end ML personalization platform for Disney+/Hulu’s unified catalog, serving real-time recommendations to 45M+ * monthly active users across web, mobile, and connected TV using Python, distributed ML systems, and model serving infrastructure. * Engineered streaming-scale recommendation data pipelines using AWS (S3, EMR), PySpark, Kafka, and SQL to transform 80M+ daily user * interactions into ML-ready features, improving data freshness by 30% and enabling consistent inputs for retrieval and ranking models. * Designed a two-stage recommendation system with collaborative filtering for candidate retrieval and LightGBM gradient boosting for * ranking, improving Precision@10 by 7% and NDCG by 5% through offline model evaluation. * Fine-tuned Llama 3 with LoRA using PyTorch and Hugging Face Transformers to generate semantic embeddings, improving content * matching by 12% and reducing cold-start recommendation errors for new titles. * Deployed ML inference services using FastAPI, Docker, and Kubernetes, achieving p99 latency under 100ms with automated model * monitoring, health checks, and rollback workflows for production reliability. * Managed machine learning lifecycle, experiment tracking, and production experiments using MLflow and A/B testing, scaling rollout from * 5% to 100% and delivering a 9% CTR improvement with enhanced user engagement.

* Developed Demand Forecasting solutions using Python, XGBoost, and LSTM Neural Networks on historical retail data from 900+ stores, * improving WMAPE accuracy by 16% compared with legacy statistical forecasting methods. * Built scalable Data Engineering workflows using PySpark, Hive, and Hadoop Data Lake architecture to process 35M+ daily store-SKU * records, enabling efficient distributed computing, dataset preparation, and batch prediction at enterprise scale. * Automated Feature Engineering, Data Quality Validation, and Workflow Orchestration using Apache Airflow, reducing refresh latency * from 5 hours to 40 minutes and eliminating 60% manual data preparation efforts. * Performed Exploratory Data Analysis (EDA), Feature Selection, Hyperparameter Optimization, and model evaluation to improve * forecasting reliability across product categories while maintaining production accuracy. * Implemented MLOps practices with MLflow, Docker, Kubernetes, and Git-based CI/CD for experiment tracking, model versioning, * automated deployment, and scalable production releases, reducing delivery cycles from 2 weeks to 4 days. * Established Production Monitoring using A/B Testing, Data Drift Detection, Bias Analysis, and performance dashboards, enabling * proactive retraining and scaling containerized ML deployments across 3 international markets while maintaining 99.5% pipeline uptime.