David Gomes

David Gomes is a software engineer at SpaceXAI who previously built database and developer-infrastructure products at SingleStore and Neon and worked on coding agents at Cursor. His work centers on making autonomous software development more flexible without surrendering control over correctness, isolation, and runtime behavior.

From distributed databases to coding agents

Gomes studied at the University of Coimbra and represented Portugal at the 2013 International Olympiad in Informatics. He joined MemSQL, later SingleStore, as an intern in 2016, progressing from full-stack development into cloud infrastructure and database systems. After returning to Portugal in 2017, he helped build SingleStore’s Portuguese engineering hub to more than 40 employees.

He joined Neon in June 2024 and contributed to Neon RLS, Neon Auth, the Neon Data API, and its Model Context Protocol server, which enables AI tools to manage database projects and execute SQL. He also worked on app.build, an open-source system for generating applications complete with databases, authentication, backend services, tests, and deployment infrastructure. Neon’s acquisition briefly made him a Databricks employee before he joined Cursor.

• Replace specialized code with agent primitives. At Cursor, Gomes reworked a large Git-worktree implementation into Markdown instructions, slash commands, and subagents. His worktree architecture supports parallel coding, multiple repositories, and moving an existing conversation into an isolated checkout while substantially reducing application-specific code.

• Make parallelism a tool for judgment. His Best-of-N model comparison workflow assigns identical tasks to different models in separate worktrees. A supervising agent evaluates their implementations, identifies substantive differences, and can combine their strongest elements.

• Measure isolation instead of assuming it. Prompted agents can drift into the primary checkout, particularly during long sessions. Gomes developed evaluations that check both whether an agent modified its assigned worktree and whether it touched forbidden files, while exploring stronger reminders, reinforcement learning, and native interfaces when firmer guarantees are necessary.

• Ground debugging in observable behavior. His analysis of agent debugging favors application instrumentation, execution traces, and textual logs over unsupported guesses. He similarly advocates AI-enabled technical interviews that test whether engineers can scrutinize generated code, catch concurrency mistakes, and remain accountable for correctness.