Nitish Agarwal
Nitish Agarwal is a Senior Engineering Manager at GoDaddy with 14+ years of experience in cloud architecture and AI. He has previously led engineering teams at Skyscanner, Expedia, and Balena, delivering high-availability systems for 80M+ users. Nitish specializes in scaling teams, optimizing distributed systems, and currently leads GoDaddy's AI transformation strategies for customer care.
he/him
nitishagar
Session
Reproducibility remains one of the most persistent challenges in scientific computing. Despite excellent tools like conda, pixi, and Jupyter, studies continue to show that a significant fraction of published computational results cannot be reproduced often due to undocumented dependencies, hidden notebook state, or fragile glue code between pipeline stages.
Meanwhile, AI agents autonomous systems that can reason, plan, and execute multi-step tasks have matured rapidly in industry settings. Frameworks like smolagents, PydanticAI, and DSPy now make it feasible to build agents that inspect environments, trace data lineage, and verify computational workflows with minimal human intervention.
This talk bridges these two worlds. Drawing on practical experience building production AI agent systems, I will present concrete design patterns for agents that serve as "reproducibility assistants" in scientific Python workflows. The talk covers three actionable areas:
(1) automated environment auditing agents that detect undeclared dependencies and version conflicts
(2) notebook-to-pipeline conversion agents that analyze Jupyter notebooks for hidden state and generate deterministic scripts
(3) result verification agents that re-execute computational steps and flag numerical divergence. A live demo will show an agent auditing a real scientific Python project end to end.