Shreya Rajpal is the co-founder and chief executive of Guardrails AI, creator of the open-source Guardrails framework and Snowglobe, a simulation platform for testing AI agents. She builds infrastructure that helps companies detect unreliable model outputs and agent failures before they reach customers.
Rajpal studied artificial intelligence at the University of Illinois Urbana-Champaign and worked on self-driving technology at Drive.ai and Apple’s Special Projects Group. She later became a founding engineer at Predibase, where she led machine-learning infrastructure. Autonomous vehicles helped shape her conviction that consequential AI systems need explicit verification and simulated testing before deployment.
In March 2023, she introduced Guardrails, an open-source package for validating language-model outputs, enforcing structured responses, and requesting corrections when results violate application requirements. The Guardrails framework turns verification-first AI development into a practical architecture: inspect each response, then accept, repair, retry, suppress, or log it.
Her technical position is unusually concrete: prompts and fine-tuning can improve behavior, but neither guarantees that an individual answer is accurate, safe, or correctly structured. Validators should match the actual risk, whether that means screening private information, executing generated code, testing SQL against a sandboxed database, or checking customer-support answers against trusted documentation. Rajpal also teaches these principles through a DeepLearning.AI course on reliable AI applications.
In August 2025, she introduced Snowglobe, extending verification from individual model responses to complete agent interactions. The simulation platform generates synthetic users and multi-turn conversations around real agents, allowing teams to test behavior without experimenting on customers.
- External grounding makes verification meaningful. Generated answers should be checked against authoritative documents, executable environments, schemas, or other application-specific evidence—not judged solely on whether they sound plausible.
- Synthetic agent evaluation requires coherent state. Useful simulations combine believable personas, mocked tools, consistent account information, and interactions whose earlier decisions affect subsequent turns.
- The simulation-to-production gap must be measured. Synthetic results deserve trust only when calibrated against human judgment and real-world outcomes. In work with Nubank engineer Aman Gupta, Rajpal described Snowglobe’s infrastructure; reported customer-service improvements and deployment acceleration belonged to Nubank.
- Strong evaluation enables high-throughput experimentation. Rajpal identifies evaluation design and rapid experimentation as essential AI-engineering skills: aligned metrics and readily generated test data let teams compare changes, isolate failures, and improve agents before release.
