Project Deep Dive (Resume Drill)

Most rounds open with 15-30 minutes of resume deep-dive before any coding. Interviewers anchor on one project from your resume, drill into the technical decisions, and probe `why this trade-off, not that one`. Requirements Standard format at the start of nearly every technical round: 5-10 min: self-introduction. Some interviewers do not introduce themselves. 15-30 min: pick one project from your resume, drill into design choices, tech stack, scale, your specific contribution. Probes: why did you choose X over Y?, what was the most challenging part?, what would you do differently?, what was the scale / latency / cost target?. Network-focused SDE screens can add how would you apply an LLM to the existing system? before coding. Backend probes can move from the data schema and indexing choices all the way to why the broader org-level infrastructure is shaped that way. For MLE rounds, the deep dive specifically targets ML projects: which models you trained, what performance metric, what improvements over baseline, how you debugged underperformance. Ads and recommendation loops can also require you to defend the project's user scale, model structure, loss, training procedure, and the mathematical reason for each choice. Notes Recent SWE / Growth / backend screens continue to spend 15-30 minutes on resume and project ownership before coding or design. Stability, online-incident handling, ownership, and the hardest project are recurring entry points. Interviewers strongly prefer projects where you clearly drove technical decisions. Pipeline-building projects get less interest than model-training projects in MLE rounds. The "why this not that" question shows up consistently — every design choice on your resume must have a defended alternative. For MLE specifically, expect the deep dive to pivot into ML 八股 mid-conversation: a project description of "I used dropout" triggers a 5-minute drill on how dropout works at inference time. A recurring failure mode is giving a fluent narrative without enough implementation detail. Prepare to defend every line of every project bullet. Bilingual interviews are common — the interviewer may switch to Chinese if your name is Chinese. Have your project pitch ready in both languages. Some interviewers re-use the same project across multiple rounds; recurring teammate dynamics suggest preparing 2-3 substantively different projects to avoid repetition. Some loops are entirely project-deep-dive driven: multiple rounds may ask no coding or behavioral questions at all, then decide based on communication efficiency, scope, and match with the team's current work. The newest SDE screens can skew even further toward the resume: expect 35-40 minutes of project and technical-detail probing inside a 45-60 minute session, leaving only 10-20 minutes for one or two coding tasks. Preparation Pick 2-3 projects, each able to fill 30 minutes of drill. Write a one-page brief per project covering: goal, scale, your contribution, key technical decisions, alternatives considered, metric impact, regrets. For each design decision in your projects, prepare a 30-second "why this not that" answer. Rehearse the project pitch in both English and Chinese. For MLE projects, also prepare the ML-knowledge follow-ups you would predictably get asked (dropout mechanics, normalization choice, loss function rationale, model selection rationale). Practice driving the conversation: open with a clear problem statement, then the design space, then your decision. Do not let the interviewer extract details one at a time — push them out proactively.