Rohan Sitaniya

At Jitterbit, I built and shipped multi-agent AI assistants that let enterprise users create integrations and publish APIs through natural language. I owned the production system behind the iPaaS AI Assistant, from LangGraph planning and Pinecone retrieval to Pydantic-validated, resumable execution.

I design AI systems around evidence, evaluation, and operational control. At Jitterbit, token-aware memory reduced LLM spend by about 40%, while retrieval improvements raised context precision by over 60% and cut latency by 30%.

At American Express, I led a five-person team building GenAI governance checks for production model reviews, cutting governance time by about 60%. I also delivered fraud embeddings, underwriting automation that reduced review cycles from days to hours, and LLM-powered credit-risk early-warning alerts.

My recent projects extend the same principle: the agent proposing a change should never control the evidence that accepts it. I build agent evaluation, guarded delivery, and research-verification systems that make failures visible before they become decisions.