At Oracle, I built LangGraph workflows for Oracle Smart Assistant, extending its retrieval pipeline with multimodal ingestion, hybrid search, and Cohere reranking. I also developed a second version of the main chat API that reduced average response time from 18 to 7 seconds.
I developed MCP servers that expose OSA capabilities to AI agents, and implemented LLM observability with OpenTelemetry tracing, token-usage tracking, and request-correlated logs. I also enhanced ground-truth testing with evaluation metrics and LLM-as-a-judge.
Earlier, I implemented software supply-chain analysis heuristics at Oracle, including a typosquatting detector that achieved 93% accuracy on the dataset used in the You Can’t Touch This paper. At Radical AI, I enhanced a Gemini-powered syllabus-generation application and contributed to an AI teaching assistant using LangChain.
