Sathya Balla

Developed ensemble models (Logistic Regression, Random Forest, XGBoost) trained on 85K+ admission records, achieving 87% accuracy in 30-day readmission prediction with 20% fewer false negatives vs legacy rules. Built end-to-end ML pipelines with scikit-learn and Apache Airflow, automating preprocessing, feature engineering, and hyperparameter tuning, reducing development time from 6 weeks to 2 weeks. Containerized inference services with Docker and deployed on Azure AKS with auto-scaling, serving 5K+ daily predictions at 99.5% uptime and <200ms response time for clinical decision support. Implemented reproducible cloud infra via Terraform and tracked 150+ model iterations in MLflow for governance and HIPAA compliance. Enabled clinical teams to prioritize 2,500+ high-risk patients per quarter, contributing to an estimated 12% reduction in preventable readmissions and $1.8M in annual savings.

Analyzed 3.2M+ transactions across 450K customers using Random Forest and XGBoost, improving demand forecast accuracy from 68% to 83% and reducing MAPE from 32% to 17%. Engineered 35+ predictive features (RFM, seasonal patterns, product co-occurrence matrices, CLV) boosting precision for high-value customer identification by 22%. Built scalable data pipelines (Redshift/S3/Kinesis) processing 50GB+ daily transactions from 15 sources at 99.7% reliability. Conducted A/B tests across 12 campaigns with 40K+ users per variant, identifying 3 campaigns with 18–25% conversion lift guiding $2M quarterly ad-spend allocation. Created Looker dashboards tracking 20+ KPIs and automated weekly Python reports, saving analysts 12 hours/week. Reduced overstock by 28% and stockouts by 19% across 200+ SKUs; improved email CTR by 31% and repeat purchase rate by 14% through better segmentation.

Led a GenAI-driven Banking Intelligence Platform; built a RAG pipeline over 500K+ banking transactions and loan records, cutting manual data retrieval time by 60%. Designed multi-agent LLM workflows in LangGraph with prompt optimization via LangSmith, boosting response accuracy from 73% to 92% across 1,200+ test cases. Deployed text-embedding-3 vectors with Pinecone semantic search, reducing hallucinations by 40% in credit underwriting. Delivered three GenAI POCs (document extraction, fraud alerts, customer service chatbot) using MCP and n8n, accelerating delivery by 25%. Powered real-time fraud detection via AWS Glue/Kinesis/RDS pipelines at 2M+ daily transactions with 99.8% uptime. Deployed inference APIs on AWS Lambda and API Gateway serving 10k+ daily requests at sub-500ms latency, reducing loan pre-approval decision time from 48 hours to 4 hours.