Sai Katteboina

• Designed and deployed production-grade RAG pipelines using Anthropic Agent Skills, MCP connectors, and LLM orchestration frameworks, reducing LLM hallucination rates by 27% across safety-critical AI applications. • Engineered low-latency ML inference pipelines leveraging PyTorch, vLLM, GPU acceleration, and high-throughput serving architectures, reducing p95 latency from 450ms to 180ms while scaling to 10K+ requests/sec. • Built automated AI red-teaming infrastructure on AWS EKS using Apache Spark and adversarial testing pipelines, improving robustness of AI safety classifiers by 35%+ against jailbreak and prompt injection attacks. • Developed reusable internal Agent Skill repositories for financial analysis, secure coding, and autonomous AI workflows, reducing repetitive engineering effort by ~15 hours per sprint. • Implemented real-time ML observability and monitoring systems using Prometheus, Kubeflow, and data quality pipelines to detect model drift and inference anomalies, enabling rapid rollback of miscalibrated classifiers. • Collaborated with AI Safety, ML Platform, and Infrastructure teams to define deployment requirements, document scalable APIs, and resolve production incidents, reducing model deployment time by 30%.

• Collaborated on a real-time fraud detection system using XGBoost, LightGBM, Apache Spark (PySpark), and ensemble ML models, processing 50K+ transactions/sec across UPI and wallet payments, reducing fraudulent TPV by 32% within 6 months post-deployment. • Built a scalable customer churn prediction pipeline for Paytm’s lending products using scikit-learn, SMOTE, and AWS EMR, achieving 0.89 AUC-ROC and enabling retention campaigns that reduced QoQ churn by 18% and saved approximately ₹4.5Cr annually. • Automated feature engineering pipelines using Apache Airflow, Redis, and distributed data workflows, reducing feature generation time from 3 days to under 4 hours while supporting 10+ production ML models. • Implemented offline-online feature skew detection and ML monitoring frameworks using Evidently AI, Kolmogorov-Smirnov (KS) tests, and data validation pipelines, reducing recommendation engine CTR degradation by 40% post-deployment. • Leveraged MLflow, Kubeflow, and Kubernetes for MLOps, experiment tracking, and model versioning, improving model redeployment reliability by 65% and accelerating iteration cycles for the payments risk team. • Developed high-performance REST APIs using FastAPI and Flask to serve real-time fraud risk scoring and churn prediction models, handling 5K requests/minute with p99 latency below 50ms and 99.95% service uptime.