James Brady

James Brady leads engineering at Elicit, an AI research assistant. His work includes setting up its original infrastructure and developing its approach to hiring AI engineers. In June 2024, he and Adam Wiggins published How to Hire AI Engineers, arguing for a combination of conventional software engineering, hands-on curiosity about language models, and a fault-first mindset.[2][3][1]

Brady describes AI application development as requiring skills comparable to distributed systems engineering, even before an application reaches large scale. Language-model calls introduce unpredictable latency, response formats, and content. In his June 2024 interview, he explained that Elicit addressed these problems at the application level through retries, timeouts, careful error handling, fallbacks, and extensive parallelization. These were established engineering techniques applied to a less predictable foundation.[1]

The team's use of strong typing made data shapes and failure cases easier to reason about. Brady explained that Python-defined types were the source of truth, with OpenAPI definitions used to generate TypeScript types for the frontend. He also described using checked exceptions within the Python codebase to make overlooked error handling harder to miss. He described this as a deliberate team practice for exposing edge cases.[1]

Availability could come at the expense of answer quality. When Wiggins raised an example of a search-provider outage producing worse results through a fallback, Brady said the team had changed search to show an error instead of serving the degraded experience. He argued that the right choice depended on the application, user, and feature. Elicit retained model fallbacks for some crucial functions, but not every task; comparable results across providers required different prompts, and backup paths could become stale without maintenance.[1]

Brady also set up Elicit's original infrastructure. The company's infrastructure-engineer role description envisages him as a sounding board and advocate for a dedicated infrastructure engineer. It separately credits Adrian “Panda” Smith with covering infrastructure part-time, including cluster bootstrapping, Cloudflare, and inference-provider work.[3][2]

Brady and Wiggins's 2024 hiring guide distinguishes AI application engineers from specialists who train models. It emphasizes proven experience with complex web applications alongside model curiosity demonstrated through self-study and serious projects. Wiggins contributed writing and helped make the team's language-model expertise accessible; Brady supplied the engineering and hiring perspective.[1]

Brady advocates interviews that simulate working together and criticizes arbitrary questions stripped of workplace context. He described coding exercises involving feature additions and bug fixes that require candidates to handle errors and edge cases. System-design discussions probe failures directly: what happens if a node dies, a network slows, or a database runs out of capacity? For Brady, these exercises also let candidates assess the team's people, tools, and working environment.[1]

In June 2024, Brady argued that centralized AI gateways could suit larger organizations with compliance or data-residency constraints. For Elicit, he considered the added abstraction a barrier to direct control over models, prompts, and rapid experimentation. His preferred focus when a new model arrived was practical: test it against existing tasks and evaluations, explore new capabilities, and ask how those capabilities could help users or advance the product.[1]

Brady also offered a personal prediction, explicitly calling it a “hot take”: “I don't think that prompt engineering is going to be a kind of durable differential skill that people will, will hold.” He expected framing the ML problem and asking the right questions to remain difficult for longer than fine-tuning a prompt's exact phrasing.[1]

Before Elicit, Brady's career centered on conventional software and engineering leadership, providing context for the skills he later brought to AI applications. He recalled leaving IBM to attend Y Combinator in 2008. He founded startups in personal knowledge management and developer tools, then brought a team to Square to build its Payroll product. At Spring, he was VP of Technology and led development of its creator-commerce platform.[1][2]

In his 2024 career account, Brady described GPT-3's 2020 release as a watershed. He considered studying machine learning formally, but in 2021 decided instead to pursue a startup focused on productizing ML-interpretability research. He approached Elicit founder Andreas Stuhlmüller as a potential advisor; in Brady's recollection, Stuhlmüller countered by suggesting that he run Elicit's engineering team. Brady said he became more excited about Elicit's application of AI to scientific problems and its alignment with safety goals than about his own startup ideas.[1][2]