TL;DR: Choose a project with real technical depth that you drove, and be ready to go several layers past the summary. Lay it out: the problem and why it mattered, the constraints, the key technical decisions and the alternatives you turned down, your specific contribution, the measurable impact, and what you would change. Expect nonstop "why" follow-ups; the round probes genuine understanding, not storytelling.
How to approach it
Treat it as a technical round, because that is what it is: at OpenAI it plays as a reverse system design, at Amazon a 45-minute Tech Talk with science Q&A. Choose a project you can defend all the way down, and open with the decision and tradeoffs, not a feature tour. Use "I" so your contribution is unmistakable on a team project.
A strong answer
Prepare the project as a layered narrative you can drill into:
- Problem and stakes. What were you solving and why did it matter (business or research impact)? One or two sentences, concrete.
- Constraints. Latency, cost, data, deadline, accuracy bar; constraints are what make the decisions interesting.
- Key technical decisions, with alternatives. This is the heart. For each major choice (model, architecture, data strategy), say what you picked, what you rejected, and why. "We used RAG over fine-tuning because the knowledge changed weekly and we needed citations" beats "we built a RAG system." Defending the alternatives you did not take is what signals depth.
- Your specific contribution. On a team project, be crisp about what you personally designed, built, and decided. Interviewers are evaluating you, not the team.
- Impact, quantified. "Cut p99 latency 40%," "improved precision from 0.7 to 0.85," "saved ~$X/month." Real numbers, honestly framed.
- Reflection. What was hard, what failed along the way, and what you would do differently. Self-awareness reads as senior.
Then prepare for the drill-down. They will ask "why" repeatedly: why that model, why not X, what was the bottleneck, how did you validate it, what broke. The round rewards someone who genuinely understands every layer of their own work and can defend the tradeoffs under pressure, and it exposes anyone narrating a project they did not really drive.
Key takeaways
- Pick a project you can defend three "why"s deep; a thin story collapses on the first follow-up.
- For every major decision, have the rejected alternative and the deciding factor ready.
- Make your personal contribution unmistakable with "I," even on team work.
- Quantify impact with numbers you can substantiate, and name one honest thing you would change.
What interviewers probe next
- "Why did you choose X over Y?" Have the comparison ready for every major decision, with the deciding factor.
- "What was the hardest part, what went wrong?" Show the real obstacle and how you diagnosed and solved it, not a polished success.
- "How did you measure success?" Tie to a metric and how you validated it (offline plus online), not a feeling.
- "What would you do differently?" A specific, credible improvement, not "nothing."
- "What exactly did you do vs the team?" Be precise and honest about your scope.
Common mistakes
- Picking a thin project you cannot defend several layers down.
- Staying at the summary level and getting exposed by the first "why."
- "We" throughout, so your actual contribution is invisible.
- No metrics, or inflated ones you cannot substantiate under questioning.
- Claiming credit for work you did not do, which collapses fast under drill-down.