CSE Roadmap

Common tier · about 4 hours · 12 parts

Part 8 of 12 · 2 min read

Answering scoping questions in an interview

How did you scope your capstone?

"One user, one job, one number. I wrote the baseline and the target before building: the share of drafted replies accepted without edits, from 45 to 70 percent, measured on 150 labelled tickets against a rules baseline. I estimated 116 hours of features and simulated the spread. Then I cut to the four core features whose 80th percentile fit the 80 hours I had." Follow-up: "What did you cut and why?" The stretch features: second language, Slack, a fine-tuned classifier, a dashboard. None of them changed the measured result.

How do you decide what counts as good enough?

"I set a usefulness threshold across quality, latency and cost before building, using what an off-the-shelf model already achieves. I remember the last mile: the reported LinkedIn example needed one month for 80 percent and four more for 95. So I pick the threshold by what the user needs and what the schedule allows." Follow-up: "What if the baseline is already good?" Then the project may not be worth building. Goals may change after evaluation.

Why use an agent here at all?

"I compare with a non-agent baseline (rules, one prompt or the human process). If it does the job at lower cost and risk, I use it. An agent has to earn its extra calls and failure modes with a measured gain." Follow-up: "How do you stage the rollout?" Crawl, walk, run: suggestions to a human first, then direct automation only for cases where the acceptance rate is high.

How do you estimate a project, and why do plans slip?

"I give each task a most-likely estimate and a spread, add a shared factor for project-wide causes, and read the 80th percentile. In the example's toy a 70-hour plan had a median finish of 78 hours and only a 54 percent chance of fitting in 80. Slips come from right-skewed task times, shared causes and the last mile." Follow-up: "What do you do when you fall behind?" I cut from the bottom of the priority list. I never cut from the evaluation harness.