At Visa, I built a Python and FastAPI incident-triage system that gives an LLM investigation tools through MCP and grounds its answers with RAG. It cut investigation time by 40% and LLM token usage by about 30%.
On OpenPathAI, I built an agent-assisted code-analysis platform with source-linked analysis and an evaluation scorecard; retrieval improvements raised Recall@10 from 0.61 to 0.86. I also published AuthDrift, an AI-agent authorization testing harness, on PyPI.

