San Francisco-based startup Patronus AI has introduced a new platform designed to prevent AI-generated inaccuracies, or “hallucinations,” in real-time, marking a significant advancement in AI safety tools. The company, which recently raised $17 million in Series A funding, claims that its platform is the first self-serve API solution focused on detecting and preventing AI failures before they reach users.
Patronus AI’s solution is built around its proprietary Lynx model, which detects inaccuracies more effectively than current leading AI models, including GPT-4. Lynx offers a dual-speed option, with real-time quick-response and deeper offline analysis, providing adaptable monitoring for a variety of industries. According to CEO Anand Kannappan, “many companies are grappling with AI failures in production, facing issues like hallucinations, security vulnerabilities, and unpredictable behavior.”
One standout feature of the platform is its customizable “judge evaluators,” which allow businesses to create plain-English rules specific to their needs, such as regulatory compliance for financial firms or patient privacy for healthcare providers. In addition, specialized tools like CopyrightCatcher and FinanceBench detect protected content and assess AI performance on financial questions, respectively, giving businesses a comprehensive safety net.
The platform’s pay-as-you-go pricing model, starting at $10 per 1,000 API calls, is designed to make advanced AI monitoring accessible to startups and smaller companies. Patronus AI has already attracted high-profile clients such as HP, AngelList, and Pearson, and has formed partnerships with major tech players, including Nvidia, MongoDB, and IBM.
As AI use continues to expand, Patronus AI’s toolset could become vital for companies aiming to meet emerging regulatory requirements, such as those outlined in President Biden’s AI executive order and the EU’s AI Act. Patronus AI emphasizes that its tools not only protect against AI errors but also help improve AI models over time, setting a new standard for ongoing AI safety and reliability.
