Chandra Pavan Kumar

Built an AI-based banking risk platform using graph analytics, knowledge graphs, and predictive models for fraud detection and transaction monitoring, improving fraud identification accuracy by 28%. Developed real-time data pipelines with AWS Glue, Kinesis, Spark, Iceberg, and Redshift to process 1M+ banking transactions for faster AI-driven risk analysis. Created AI agents with LangGraph, CrewAI, AutoGen, and MCP to support fraud/AML/credit workflows, reducing investigation time by 9% using automated case analysis and intelligent retrieval. Implemented GraphRAG with Neo4j vector databases and LLMs to provide risk insights and improve alert prioritization, increasing investigator productivity by 15%. Fine-tuned Llama and Mistral models with QLoRA/PEFT and hyperparameter tuning to optimize performance for risk prediction and fraud/credit tasks. Used MLflow, Kubeflow, Docker, Kubernetes, Terraform, and AWS services to manage deployment, monitoring, and governance, improving production mo

Built an AI-based banking risk platform using graph analytics and knowledge graphs to support fraud detection and transaction monitoring, improving fraud identification accuracy by 28%. Developed real-time data pipelines with AWS Glue, Kinesis, Spark, Iceberg, and Redshift to process 1M+ banking transactions and customer activity data for faster risk analysis. Created AI agents with LangGraph, CrewAI, AutoGen, and MCP to streamline fraud/AML/credit workflows and reduce investigation time by 9% through automated case analysis and information retrieval. Applied GraphRAG with Neo4j vector databases and language models to generate risk insights, improve alert prioritization, and increase investigator productivity by 15%. Fine-tuned Llama and Mistral models using QLoRA/PEFT and hyperparameter tuning, and used MLflow, Kubeflow, Docker, Kubernetes, Terraform, and AWS services to improve production model reliability by 22%.

Designed a real-time rider-driver matching system using embedding models and H3 geospatial indexing to improve dispatch accuracy and reduce pickup latency by 18% at Uber scale. Built streaming feature pipelines with Apache Kafka and Apache Flink for real-time ETA prediction and dynamic pricing signals, improving surge pricing accuracy by 22%. Enhanced marketplace ranking and driver incentive models using gradient-boosted learning, improving trip conversion rate by 16% through large-scale A/B testing (reported production metrics included Precision@K=0.87, Recall=0.81, F1=0.84, ROC-AUC=0.90). Implemented scalable ingestion and transformation pipelines for marketplace systems, standardizing event-driven architecture and improving indexing throughput by 21%. Deployed production inference services with Docker and Kubernetes and CI/CD pipelines, supporting real-time predictions for 2.1M+ marketplace records. Refactored legacy analytics into microservices and built Python evaluation framework