Technical leadership
A senior AI/ML engineer interview on leading model and LLM work without formal authority, from launch bars and stakeholder calls to incidents, budgets, and growing the team.
These prompts are original practice. They are not questions from a named employer or a live hiring loop.
1.You are the tech lead for a new support-reply assistant built on a large language model (LLM). Three engineers each have a prototype using a different approach. Leadership wants one plan by next month. What do you do first?
2.Your design doc for a new ranking model is 30 pages long. Reviewers keep skimming it and approving without real comments. What is the best change?
3.You are drafting the technical direction for your team's retrieval system for the next year. What should the doc make clearest?
4.In a design doc for a retrieval-augmented generation (RAG) feature, you list only the approach you favor. A reviewer asks what else you considered. What is the best response?
5.Your team keeps building one-off prompt pipelines for each product request. You want to propose a shared platform. What makes the proposal most likely to be accepted?
6.A staff engineer on another team says your design doc for a model gateway is fine but too vague to build from. What is the most useful next step?
7.You plan to move the team's model serving to a new framework. The doc is ready. Who should review it before you circulate it widely?
8.Your doc recommends a hosted model API for a summarization feature. Leadership asks for one sentence on why. Which is best?
9.In a design review for a new embedding pipeline, a senior engineer strongly objects to your choice of index. The rest of the room is silent. What do you do?
10.During review, a junior engineer points out that your evaluation set leaks training examples. You had not noticed. What is the best response?
11.Your design reviews keep running long because two engineers argue about framework choices that do not affect the outcome. What do you change?
12.After a design review, you decided against an engineer's proposal to fine-tune a model. They are still unhappy. What should you do?
13.A reviewer asks for a load test before approving your new inference service. The launch is in four days and you think the test is not needed. What is the strongest move?
14.You notice that in your team's design reviews, the same two senior people do almost all the talking. What is a good fix?
15.A principal engineer from another org reviews your design for a shared feature store and says it will not scale. You have data showing it will at the projected load. What do you do?
16.Your team adopted a rule that every model change needs a design review. Small prompt tweaks now wait a week. What do you propose?
17.Product wants to know whether to fine-tune an open model or use a hosted model for a contract-review feature. You do not have enough data to be sure. What do you do?
18.Legal says customer documents cannot leave your own cloud account. Your team prefers a hosted model from a vendor. What should guide the decision?
19.Six months ago you chose a hosted model for search answers. Usage has grown tenfold and costs now far exceed the original estimate. What is the right next step?
20.Two embedding models score within one point of each other on your evaluation, which is inside the run-to-run noise. The team has argued for two weeks. What do you do?
21.You need to pick a vector database. One option is mature but slow. Another is fast but young. What should you do before deciding?
22.Your manager asks you to choose between two model vendors by Friday. You think you need two more weeks. What is the best response?
23.A year ago your team built its own model training platform. A newer managed service now does most of the same job. What should you do?
24.You are leading a large language model (LLM) feature from prototype to production. The prototype works in a notebook. What should the first production milestone be?
25.Your team has a quarter to ship a document question-answering feature. Halfway through, it is clear you cannot do all five planned document types. What do you do?
26.A director asks your team to add voice input to the assistant this quarter. Your team is already at capacity on core quality work. What do you do?
27.You are planning a multi-quarter project to replace a rules-based fraud system with a model. How should you sequence it?
28.An engineer wants to add a reranking model, a query rewriter, and a cache to the first version of a search assistant. What do you advise?
29.Sales has promised a customer that your summarization feature will support twelve languages next month. Your evaluation covers three. What do you do first?
30.You are asked to estimate how long it will take to raise answer accuracy from 80 to 90 percent. What is the honest way to answer?
31.Midway through a project to add tool use to your assistant, a new model release makes half your planned work unneeded. What do you do?
32.A product manager asks your team to build something that would only benefit one small customer. You think it is a poor fit. What is the best way to say no?
33.You are setting the evaluation standard your team must meet before any large language model (LLM) feature launches. What should the standard include?
34.A team asks to launch a feature that misses the agreed evaluation bar by two points. Their deadline is tomorrow. What do you do?
35.Your team's evaluation set was built two years ago. Users now ask very different questions. What should you do?
36.Engineers on your team keep tuning prompts until the evaluation score goes up, then the gains vanish in production. What should you change?
37.A product lead wants to skip your safety evaluation for an internal-only tool. What is the strongest response?
38.Your team uses a large language model (LLM) as a judge to grade answers. A senior engineer doubts its grades. What is the best way to settle it?
39.Different teams in your org use their own evaluation scripts, and their scores cannot be compared. You want a shared standard. What is a good first step?
40.Your launch checklist says the model must be evaluated for fairness across user groups. A team says they have no group labels. What do you do?
41.Leadership asks for one number that shows whether the assistant is good enough. What should you give them?
42.You find that a feature you approved last month was evaluated on data that excluded long documents, which are a third of real traffic. What do you do?
43.Your agent demo impressed leadership. In a pilot with real users, it fails on about a third of tasks. Leadership expects a launch next month. What do you do first?
44.An engineer on your team wants to run the quarterly demo using hand-picked inputs that show the model at its best. What do you advise?
45.Your assistant scored well on the internal test set, but users rate it poorly. What is the most useful first step?
46.A vice president saw a demo and announced the feature to customers before your team agreed it was ready. What should you do?
47.In the demo, your model answered in two seconds. Under real traffic, answers take twelve. What should the team have done before the demo?
48.Your prototype used a small set of clean documents. Production documents are scanned, messy, and much longer. How should you plan the next phase?
49.After a strong demo, leadership wants the feature in all regions at once. You know it has only been tested in English. What do you recommend?
50.Your training pipeline is held together by scripts only one engineer understands. You want time to fix it. How do you make the case to your director?
51.Your team has many old prompt versions, unused features, and test sets nobody runs. Engineers keep tripping over them. What is a sensible plan?
52.You inherited a recommendation model trained on a feature that another team is about to deprecate. What do you do first?
53.Your team took a shortcut to launch: the model's thresholds are hard-coded and tuned by hand. Six months later, every change needs a full retest. What should you propose?
54.During planning, your team lists twenty pieces of technical debt. How should you decide which to fix?
55.Your director turned down a debt project you proposed. You still think it matters. What do you do?
56.You lead a migration of eight teams from an old model-serving stack to a new one. Two teams have not started after a month. What do you do?
57.You plan to deprecate an old embedding model that twelve services use. How should you begin?
58.Halfway through moving teams to a new prompt library, you find a design flaw that will force a breaking change. What do you do?
59.A team refuses to move to the shared evaluation harness because their metrics are special. What is the best response?
60.Your team is moving models from one cloud GPU provider to another. How do you know the move is safe for each model?
61.The migration you lead is three months in, and the last 10 percent of callers are the hardest. Leadership is losing interest. What do you do?
62.Your migration requires each team to rewrite their inference calls. How can you cut the work for them the most?
63.Your team keeps shipping prompt changes that break other features. Code review catches some, but not all. What should you do?
64.In code review, you keep leaving the same comment about logging model inputs and outputs. What is the best fix?
65.A strong engineer on your team ships fast but often skips tests for model changes. Others are starting to copy them. What do you do?
66.Your team wants a standard for when a model change needs a staged rollout. What should the standard be based on?
67.Code reviews on your team take days, and people are frustrated. What change will help most?
68.A model went live with a data preprocessing bug that a simple test would have caught. In the follow-up, what should you push for?
69.A mid-level engineer wants to lead the next model launch. You think they are close but not quite ready. What do you do?
70.An engineer you mentor keeps bringing you problems and asking what to do. What is the best way to help them grow?
71.A research-minded engineer wants another month to try a new training method. The feature is due in two weeks, and the current model meets the bar. What do you do?
72.Two engineers on your team always get the hardest model problems, while others get routine work. What should you change?
73.A junior engineer proposes a model architecture you think will not work, but trying it is cheap. What do you do?
74.Your research scientists publish strong results, but your product engineers struggle to use them. What should you do?
75.An engineer on your team is great at modeling but avoids writing design docs. They want to be promoted to senior. What feedback helps most?
76.You delegated the evaluation harness to an engineer. They chose a design you would not have picked, but it works. What do you do?
77.The team's scientists say shipping pressure leaves no time for careful experiments. Product says the team ships too slowly. What do you do?
78.A new hire on your team is quiet and has not asked any questions in three weeks. What should you do?
79.Your team's model launch is two weeks away, and you now think it will be three weeks late. What do you do first?
80.Your project status has been green for weeks, but you now realize the main risk was never tested. What should you do?
81.Your data labeling vendor is behind schedule, putting model training at risk. What do you do?
82.A stakeholder gets weekly updates on your model project but still seems surprised by delays. What should you change?
83.Your manager asks you to keep a slipping project's status green until after a leadership review. What do you do?
84.Your model project is on time, but quality will likely miss the agreed bar. Stakeholders have not asked. What should you do?
85.Your chat assistant told a user something harmful, and a screenshot is spreading on social media. You are the tech lead. What do you do first?
86.After an incident where your model gave harmful advice, an engineer says the user tricked it, so it is not a real problem. What do you say?
87.You are running the follow-up review after your model leaked parts of a private document in a reply. What should the review focus on?
88.During an incident, three senior engineers are trying different fixes in production at the same time. You are the lead. What do you do?
89.After an incident, the team lists ten action items. Three months later, none are done. What should you change?
90.A model that flags harmful content starts letting more through after a data change. You learn about it from a user report. What should the follow-up add?
91.Your director wants to know who caused the incident where the assistant gave wrong medical information. What do you say?
92.An engineer on your team caused an outage by pushing a model config with a typo. They are upset. What do you do?
93.Your team and another team both need most of the shared GPU budget next quarter. Neither will back down. What do you do?
94.Your training jobs keep getting preempted by another team's experiments on the shared cluster. What is the best first step?
95.Product wants your assistant to answer more questions. Trust and safety wants it to refuse more. You sit between them. What do you do?
96.Your model depends on features owned by a data team. They changed a feature without telling you, and your model degraded. What do you do?
97.Design wants the assistant to sound warm and casual. Legal wants careful, formal answers. Both ask you to pick. What do you do?
98.Another team asks you to add a feature to your model service that only they need. It would add risk for every other caller. What do you do?
99.Finance asks your team to cut GPU spend by 30 percent next quarter. What is the best first step?
100.Your team and a partner team disagree on which embedding model to standardize on. Both have good reasons. The decision is blocking work. What do you do?
101.Your manager wants to launch a model you think is not safe yet. What is the best way to disagree?
102.You disagreed with your director about a model vendor, but they chose the other one. What should you do?
103.You want other teams to adopt your team's safety filter for model outputs. You have no authority over them. What works best?
104.A more senior engineer insists on a training approach you believe is wrong. You have data. What do you do?
105.Every model change on your team waits for your review, and you have become the bottleneck. What should you do?
106.You think your team should spend a quarter on evaluation tooling instead of new features. Your manager disagrees. What should you do?
107.You are asked to take over a struggling model team for a quarter. The team is tired and wary of you. What do you do first?
108.You notice that you write most of the design docs and lead most of the meetings on your team. What does a staff engineer do?
109.A senior leader asks you in a meeting whether your model can handle a new use case. You are not sure. What do you say?