Sharad Babar.

NVIDIA

Apr 2025 - Present

Data Scientist

New York, Hybrid

• Optimized distributed LLM training pipelines for 30M+ parameter models using NVIDIA NeMo and PyTorch on A100/H100 DGX Cloud clusters, improving throughput by 32% and cutting compute costs by 18%.
• Built scalable preprocessing and tokenization pipelines with Python, SQL, and Apache Spark across 50+ TB of unstructured text data, improving downstream model convergence stability by 22%.
• Fine-tuned transformer-based architectures with Hugging Face Transformers and PyTorch Lightning, improving task-specific accuracy by 15-25% across benchmark datasets.
• Developed LLM evaluation and monitoring frameworks with Python and MLflow to track hallucination rates, perplexity, and latency across 100+ training runs.
• Deployed end-to-end MLOps pipelines on AWS with automated CI/CD for retraining and model versioning, reducing deployment time from 5 days to under 24 hours.

Accenture

Jun 2020 - Jul 2023

Data Scientist

India

• Engineered a real-time fraud detection system using Python, PySpark, and Scikit-learn, detecting 14M+ fraudulent transactions annually and improving model accuracy by 24%.
• Architected Spark, Kafka, and SQL ingestion pipelines for 2M+ daily financial transactions, enabling near real-time fraud scoring and reducing decision latency by 35%.
• Built feature engineering frameworks with Pandas, NumPy, and Spark SQL that improved model precision and recall by 20%+ in high-risk fraud segments.
• Implemented Logistic Regression, Random Forest, XGBoost, and TensorFlow-based deep learning models, improving production AUC from 0.82 to 0.91.
• Developed Power BI dashboards and KPI monitoring for 100+ enterprise banking clients, giving stakeholders real-time visibility into fraud trends and alert volumes.