At Cummins Inc, I designed and deployed a Databricks MCP server that lets Microsoft Copilot access enterprise data securely. I also built a LangGraph agent that propagates user-level permissions across Databricks, Snowflake, and Palantir.
I built a production RAG retrieval system and used MLflow to evaluate chunking strategies and prompts. After tracing a 30% hallucination rate to character-based chunking, I introduced token-based semantic chunking and citation verification, reducing unsupported claims to under 5%.
At Cummins, I developed a FastAPI service to turn pricing rules into executable Python constraints, then replaced the LLM with a deterministic script that saved over $500 per month and reduced manual processing time by 95%. I also migrated pricing analytics from Pandas to PySpark, reducing execution time by 72%.
Earlier, I built data pipelines and dashboards at Capgemini, and developed data-labeling and research tools at Luddy School of Informatics and Kelly School of Business. In my projects, I built a production RAG API and a multi-agent stock analysis platform with automated rebalancing recommendations.
