← Forward Deployed Engineer
Forward Deployed Engineer

A Working Demo Is Not the Finish Line

3 Nov 2026·3 min read

OpenAI’s own careers page states that FDEs are measured by "production adoption, measurable workflow impact, and eval-driven feedback" — not by whether the demo went well. A demo proves an idea can work. Production proves it keeps working when no one is watching.

What a demo almost never has to survive

  • Real data volume, not a curated sample.
  • The customer’s actual uptime and latency expectations, not best-effort.
  • Nobody standing by to restart it when something breaks at 2am.
Why this gap is so easy to underestimateA demo and a production system can look identical on screen. The difference is entirely in what happens when inputs are unexpected, load is high, or nobody is present to notice a silent failure — none of which is visible in a 15-minute walkthrough.
Do this in 60 secondsTake any small script you have that "just runs" when you run it. List three things that would make it unreliable if it ran unattended, daily, for a month — think about what happens to logs, what happens if an input is missing, and what happens if it runs twice by accident.
Today’s AI angleThis gap is exactly why evals matter: an LLM-based demo answering five curated questions well tells you almost nothing about how it will handle the thousands of real, messy questions production actually brings.
Interview prepYour demo worked perfectly. The customer wants it live next week. What do you check before that happens? I check what the demo never had to survive: real data volume instead of a curated sample, the customer's actual uptime expectations, and what happens with nobody watching if it fails at 2am. OpenAI's own framing is the right one here — a demo proves an idea can work, production proves it keeps working, and those are different bars.
A Working Demo Is Not the Finish Line · Ali Huzaifa