OpenAI FDE Luis Velasco: production AI starts with evals, not code
In a ShiftMag interview, OpenAI forward deployed engineer Luis Velasco says his job is closing the "deployment gap" of data, procedures and tools, and that evals, permissions and system design matter more than writing code.
ShiftMag sat down with Luis Velasco, a Forward Deployed Engineer at OpenAI and former Googler, at the Shift conference in Zadar. He sums up the job as helping companies put AI to work in real environments, turning demos into production systems.
Velasco says the hard part begins when a demo meets production. OpenAI's FDEs lean heavily on eval-driven development: building representative eval sets with easy, medium and edge cases, clearing them before launch, and rerunning them to catch regressions when a customer upgrades models. He describes the customer's data, procedures and tools as the "deployment gap" that FDEs exist to close.
On security, he argues agents should get the same scoped access rules as employees, seeing only what their workflow needs. And on careers, he says code is now cheap to produce, so the valuable output is the harness around the model: intent, guardrails and constraints. "System design taste will never go away," he says.
For engineers eyeing FDE roles at AI labs, the interview is a useful snapshot of what the work actually rewards: evals, integration and production judgment rather than raw coding speed.
Source: shiftmag.dev