Johann Schleier-Smith

Johann Schleier-Smith is the technical lead for AI at Temporal Technologies, a computer scientist, and a serial founder working to make AI-generated software and autonomous applications dependable. His career spans large-scale social networking, foundational research on serverless computing, AI-assisted database administration, and the infrastructure needed to keep agents running reliably in production.

From social networks to reliable AI

Schleier-Smith studied physics and mathematics at Harvard and pursued doctoral work in physics at Stanford before leaving in 2004 to co-found the social-networking company Tagged and if(we) with Greg Tseng. As chief technology officer, he helped build a business encompassing the Tagged and hi5 brands that ultimately reached more than 300 million people and joined The Meet Group in 2017.

In 2015, he left his executive role for graduate study at Berkeley, remaining on if(we)’s board. He completed a computer science master’s degree in 2016, began doctoral research in 2017, and earned a PhD with a 2022 dissertation on serverless cloud computing.

At Berkeley, Schleier-Smith investigated how programmers might use distributed infrastructure without individually managing servers. He co-authored an influential analysis of serverless computing’s unresolved challenges, contributed to Cloudburst, a platform for stateful serverless functions, and led research on a transactional serverless file system.

He subsequently founded Crystal DBA, which applied AI to PostgreSQL administration, troubleshooting, and performance. The Crystal DBA team later joined Temporal, where he develops reliability infrastructure for AI applications. His Workflow Streams demonstrations show analytics and voice agents streaming responses through durable workflows that can reconnect without losing execution state.

• Make high-assurance software affordable. Schleier-Smith’s zero-bug software vision adapts aerospace practices—detailed specifications, defensive programming, independent verification, and redundant implementations—to AI-assisted development. His proposition is economic: cheaper generation and testing could bring stronger assurance to ordinary software, without guaranteeing the elimination of every defect.

• Verify the specification, not merely the output. His demonstration of formal verification uses Dafny to embed explicit requirements alongside executable code and reject an introduced error before generated Python runs. The guarantee extends only to properties the specification actually captures.

• Keep generation and scrutiny independent. He advocates separate prompts for implementation and testing, explicit risk analysis, written safety cases, and potentially different model providers. Trusted libraries, modular architecture, and compositional reasoning also reduce the verification burden imposed by limited model context.

• Separate operational durability from correctness. His work on durable execution for AI agents addresses crashes, network interruptions, long-running tasks, and reconnectable streams. These guarantees preserve an application’s progress; they do not independently establish that an agent’s decisions or generated code are correct.