seniorReportedcurated
Tell me about a time you chose a simpler model or architecture over the more sophisticated option.
30-second answer
- The decision was to ship the smallest thing that could be measured: a single prompted classifier with a small eval set and a confidence threshold that routed low-confidence items to humans.
- That rejected the planner and multi-agent proposal because classifying inbound mail into eight categories does not need role decomposition, and every added layer is another thing to debug.
- The simpler version gave up multi-intent handling, so roughly 9% of traffic went to review instead of being split automatically, and that was a named cost rather than a surprise.
- It still reached 94% routing accuracy at about a third of the latency and a fraction of the token cost, and shipped three weeks earlier.
- The process change was to require a measured gap before adding an orchestration layer, and to treat multi-intent as a roadmap item rather than a launch blocker.
- The limit is that the simpler classifier cannot grow into multi-intent routing, so the review queue stays a standing cost until a splitter is built.
Likely follow-ups
Open a question to rehearse the answer.
When would the multi-agent version have been right?
- The multi-agent version would have been right if multi-intent splitting were the product requirement, because that is a task the single classifier cannot represent.
- It would also have been right if stages needed genuinely different tools and acceptance criteria, which is the only condition that makes role decomposition more than decoration.
- It would have been justified once the review queue was dominated by items a planner could route, so the measured gap paid for the added cost.
How did you measure the cost and latency difference?
- I ran the same labelled set through both systems and compared routing accuracy, reviewer rate, p50 and p95 latency, and tokens per request.
- I converted tokens to cost per thousand requests, because token counts alone do not show what the difference means at traffic.
- I kept the simple baseline runnable, so the comparison could be repeated when the multi-agent version changed rather than argued from one snapshot.
Red flags
- Frames simplicity as laziness rather than a measured decision
- Has no numbers for what the simpler version gave up
Concepts this tests
Tags
Source: engineersofai.com
More in Behavioural & project deep-dives
- Walk me through an AI feature you shipped end to end and your exact role. What decisions were yours?
- Tell me about a time a model surprised you in production. How did you find out, and what did you change?
- Describe a time you disagreed with a product or engineering stakeholder about using an agent, fine-tuning, or RAG.