Bharath Kumar

At Marlo Tech Asia, I built and shipped a production RAG and agent platform that gave operations teams natural-language access to vessel activity and history across about 10,000 emails per day. Experiments with hybrid retrieval brought the labeled evaluation set to 86% Recall@10.

I developed LangGraph agents with search, knowledge-graph, and task tools, including Jira integration through MCP. I also built a RAGAS-based evaluation framework and regression tests to track retrieval and answer quality.

For an enterprise data platform, I designed a hierarchical multi-agent system on AWS Bedrock for SQL generation, Athena retrieval, Python analysis, and visualization. I added read-only execution and permission controls, then built FastAPI services and CI/CD pipelines on Amazon EKS and GitHub Actions.

At Thinq24 Innovations Private Limited, I fine-tuned BERT for maritime email classification and built a vessel risk-scoring pipeline with PySpark and Databricks. The classification system reached 0.85 macro-F1, while the XGBoost risk model reached 0.75 AUC.