"Walk me through an AI project you've built. What problem did it solve, how did you design it, and what would you change?"

The single most common opener in every AI/Applied Engineer phone screen. This session is entirely about answering it — practice out loud, not just in your head, until it comes out in 60-90 seconds unprompted.

Miss one and the follow-up lands exactly where you went silent.

1. Problem — a named constraint, not just what the output does
2. Architecture — deliberate layer design with named roles
3. Trade-offs — alternatives compared, the decision quantified
4. Reliability — a named mechanism for what happens when something breaks
5. Security — a named threat and a named control

"I built a Researcher Agent — a system that answers multi-step research questions using a 5-layer architecture: gateway, orchestrator, LLM provider, cache, and observability. The constraint I was solving for was cost-and-reliability: a single LLM call is cheap but unreliable for complex questions, so I implemented a ReAct loop capped at 5 steps with retry on every call. I benchmarked two model options and chose the mid-tier one — 40% quality improvement justified a 3x cost increase for this use case. My eval harness gives me a 7/10 accuracy score across 3 test inputs; the known failure mode is ambiguous temporal queries — 'what happened recently' with no date anchor. The main thing I'd change is adding streaming output from the start — the synchronous loop takes 5-8 seconds and that's a UX problem. On security: my STRIDE model flagged prompt injection as the top risk, so I implemented query length and pattern validation in the gateway."

Every sentence names a number, a layer, or a mechanism — nothing here is a vague gesture at "good engineering practice."