Rajiv Shah is an agentic AI engineer at OpenHands who builds and teaches the infrastructure that makes autonomous coding agents reliable in enterprise environments. His career spans industrial data science, automated machine learning, open-source models, retrieval systems, and practical evaluation of AI agents.
Shah served in the U.S. Coast Guard and earned a PhD and a law degree from the University of Illinois. He taught himself data science using R and Python before working at State Farm and Caterpillar on actuarial modeling, supply chains, sensor data, and cybersecurity. He is a named inventor on a State Farm patent for detecting anomalous computer behavior.
At DataRobot, Shah helped organizations apply automated machine learning; subsequent roles took him through Snorkel AI, Hugging Face, and Snowflake. His open-source work includes LLM evaluation notebooks and models and datasets published through his Hugging Face profile.
As Contextual AI’s chief evangelist, Shah concentrated on enterprise retrieval. At AI Engineer World’s Fair 2025, he and Nina Lopatina led a workshop on production-ready retrieval agents, addressing document extraction, hybrid search, reranking, grounded responses, evaluation, and organizational controls.
In 2026, Shah joined OpenHands, expanding his focus from retrieval to autonomous software engineering and the systems that constrain, execute, and assess agents.
- Agentic harness engineering: Useful agents require more than a capable model: routing, tools, memory, orchestration, verification, and safe execution determine whether they finish meaningful work. Shah’s hands-on OpenHands course turns those architectural concerns into practical exercises.
- Outer-loop coding agents: Autonomous background agents can investigate vulnerabilities, propose fixes, modernize legacy code, and test infrastructure changes. Enterprise deployment demands sandboxing, permission boundaries, network controls, and governance alongside autonomy.
- Managed retrieval-augmented generation: Production retrieval depends on accurate document parsing, hybrid search, reranking, grounded generation, and source attribution. Shah emphasizes that misread tables, missing document hierarchy, and weak permissions can undermine an application before a model generates its answer.
- Measurable agent-skill evaluation: In his framework for evaluating agent skills, Shah advocates bounded tasks, deterministic pass-or-fail checks, and comparison against a no-skill baseline. Because guidance can help one model and hinder another, he treats agent improvements as measurable hypotheses.
