Patronus AI

Patronus AI is an automated platform for evaluating, monitoring, and guardrailing LLM applications and AI agents, built so teams can catch failures like hallucinations, prompt injection, and unsafe output before they reach users. Founded in 2023 in San Francisco by former Meta FAIR researchers Anand Kannappan and Rebecca Qian, it began as a self serve evaluation and security API and has since expanded toward agent simulation. Its central idea is that good evaluation is not just a safety net but a way to systematically improve AI products, and it backs that with proprietary models trained by its own research team.

At the core are research grade evaluators, most notably Lynx, a hallucination detection model that the company reports beats GPT-4o at catching inaccuracies in retrieval augmented generation, and GLIDER, a general purpose small model judge. These come in small variants for low latency real time guardrails and larger variants for deeper offline analysis. Developers embed the Patronus API directly in their code through a language agnostic interface and Python SDK, configure custom LLM judges by describing criteria in plain English, and get back not just a verdict but the exact span of text where a problem occurs. Specialized tools include CopyrightCatcher for reproduced protected content and the FinanceBench benchmark, and the platform adheres to OWASP and NIST standards.

Beyond guardrails, Patronus provides a full platform around evaluation. Patronus Experiments runs side by side A/B tests of prompts, models, and RAG configurations, while production logs and traces feed a dashboard for tracking results and comparing performance over time. Percival is an agent debugger that automatically detects more than twenty failure modes in agentic execution traces and suggests prompt and workflow fixes. In late 2025 the company moved into simulation with what it calls Digital World Models, reinforcement learning environments, and generative simulators that stress test agents against adversarial and chaotic scenarios, a direction funded by a $50M round in 2026.

Patronus is offered as a usage based, pay as you go service. New users sign up, create an API key, and start with five dollars in free credits, then pay per API call, with published rates of around ten dollars per thousand calls for smaller evaluators and twenty dollars per thousand for larger ones. Because cost scales linearly with API traffic, high volume production guardrailing needs careful cost modeling. Enterprise plans add higher rate limits, custom evaluation models, webhooks, and professional services. Customers include AngelList, Pearson, HP, and Fortune 500 firms in regulated industries.