Practice

72 questions

• Statistics

  Revenue management asks for "the 95% interval for passengers on flight ZX104 next Friday". A colleague returns the 95% confidence interval of the mean passenger count over the last 52 Fridays.
  senior8 min

• Statistics

  Ancillary attach rate (share of passengers who buy a paid seat or bag) rose in every market segment this quarter, yet the company-wide attach rate fell. The CFO thinks the data are broken.
  senior8 min

• Statistics

  You need a 95% interval for average ancillary revenue per passenger. Most passengers spend nothing and a few spend a lot (group upgrades, excess baggage).
  senior10 min

• SQL and Python

  You have 10 million booking events (customerid, bookedat, flightno). For each booking, flag whether the same customer made another booking within the previous 24 hours.
  senior15 min

• SQL and Python

  You are building a training table for a churn model: one row per customer per monthly cutoff, with features such as loyalty tier and number of bookings in the previous 90 days.
  senior12 min

• SQL and Python

  A nightly pandas feature job that used to take 20 minutes now takes 3 hours and occasionally runs out of memory. How do you diagnose and fix it?
  senior12 min

• Analytics

  Product says the new loyalty app is working: "Members who installed the app last year have 30% higher repeat-booking rates than members who did not." How do you evaluate the claim, and what analysis…
  senior12 min

• Analytics

  Finance reports monthly ancillary revenue 3% lower than the product analytics dashboard, and both teams say their number is correct. You are asked to resolve it. How?
  senior12 min

• Analytics

  You must set up automatic alerts on 200 daily KPIs (bookings, revenue, load factor by market and channel).
  senior12 min

• Classical ML

  In a gradient-boosted model of booking cancellations, SHAP says "days to departure" is the top driver and that higher price lowers cancellation.
  senior12 min

• Classical ML

  A no-show model trained on a dataset where no-shows were undersampled to 50/50 is now used to compute expected no-shows per flight for overbooking.
  senior12 min

• Forecasting

  You build one global LightGBM model to forecast daily passengers for 300 routes, 14 days ahead. Features include lag1, lag7 and a 7-day rolling mean. Backtest error is excellent.
  senior12 min

• Forecasting

  You are asked to forecast passenger demand for a route to support a decision on adding a second daily flight. Historical passengers are often at capacity on the existing flight.
  senior12 min

• Forecasting

  Airline catering must load meals per flight. Running out of meals costs about 300 per missing meal (compensation, complaints), while an unused meal costs about 100 (hypothetical figures).
  senior12 min

• DL, NLP and LLMs

  An airline wants an assistant that answers agents' questions about fare rules, baggage policy and rebooking procedures. Policies change monthly.
  senior15 min

• DL, NLP and LLMs

  Explain how LoRA works. For one 4096 x 4096 weight matrix with rank 8, how many trainable parameters does LoRA add, and what are the practical trade-offs compared with full fine-tuning?
  senior12 min

• RAG and agents

  Users of an internal HR-policy assistant report confident but wrong answers. Product wants to swap in a bigger LLM. How do you find out what is actually wrong before changing anything?
  senior15 min

• RAG and agents

  You are building an email assistant agent that reads incoming customer emails, looks up bookings and can send replies and issue travel vouchers.
  senior15 min

• RAG and agents

  A team proposes an autonomous agent to process flight-disruption refund requests end to end. The process has clear eligibility rules but messy inputs (free-text emails, attachments).
  senior12 min

• Experimentation and causality

  An experiment configured for a 50/50 split ends with 50,600 users in control and 49,400 in treatment. Treatment shows a significant +2% revenue lift. Do you trust it?
  senior10 min

• Experimentation and causality

  Revenue management introduced a new overbooking policy on 12 routes in March; the other 40 routes kept the old policy. Routes were chosen by managers, not at random.
  senior15 min

• Experimentation and causality

  A telecom company sends retention discounts to the 10% of customers with the highest churn-model scores.
  senior12 min

• Experimentation and causality

  A redesigned seat map lifts paid seat selection by 6% in week 1, 3% in week 2 and 2% in week 3. Separately, product complains that tests take too long.
  senior12 min

• Production and system design

  Design a churn-prediction system for a subscription software company with 2 million customers. The retention team contacts at-risk customers weekly.
  senior40 min

• Production and system design

  Design a daily demand-forecasting system for a grocery chain: 500 stores, 20,000 SKUs, forecasts 1 to 14 days ahead feeding automatic replenishment.
  senior40 min

• Production and system design

  Describe how you would deploy a new version of a model (or a new prompt for an LLM feature) safely, and how you would decide to roll back. What must be in place before the first deployment?
  senior15 min

• Business judgment

  Before anyone builds it, the head of marketing asks you: "Is a churn model worth it for us?" The company has 1 million subscribers and about 2% monthly churn.
  senior15 min

• Business judgment

  The chief commercial officer asks: "Should we launch a cheaper basic fare without a checked bag on our domestic routes?" You have 20 minutes. How do you structure your answer?
  senior20 min

• Communication

  You spent three weeks analysing why on-time performance fell on regional routes. The COO has five minutes and will read one page.
  senior15 min

• Communication

  Two days before a promised launch, you discover that the demand model's validation used a feature that leaks future information.
  senior12 min

• Communication

  Tell me about a time you disagreed with a senior stakeholder about a technical or analytical decision and had to influence them without authority. (Behavioural: answer using STAR.)
  senior10 min

• Statistics

  An airline flags the 10% of routes with the worst on-time performance (OTP) in Q1 and sends them a punctuality programme. In Q2 those routes improve from an average of 71% to 77% OTP.
  senior12 min

• Statistics

  An A/B test randomises logged-in users, but the analyst computes the standard error of revenue per session with every session treated as an independent row: 10,000 users, about 5 sessions each, a…
  senior12 min

• SQL and Python

  You have bookings(bookingid, route, bookedat) and a table of fare snapshots faresnapshot(route, snapts, fare) that is appended whenever the price engine publishes a new fare, a few times per day per…
  senior15 min

• SQL and Python

  A web-search event log events(customerid, ts) has one row per search. Define a session as a run of events where no gap between consecutive events of the same customer exceeds 30 minutes.
  senior12 min

• Analytics

  A funnel dashboard reports search-to-purchase conversion at session level (hypothetical figures): mobile 60,000 sessions and 3,000 purchases (5.0%), desktop 40,000 sessions and 3,000 purchases…
  senior12 min

• Classical ML

  A team tunes a gradient-boosting model with a random search of 200 configurations on one validation set and reports the best validation AUC (0.83) as the expected production performance.
  senior12 min

• Experimentation and causality

  A new in-app 'price alert' feature is rolled out in an A/B test: 50,000 users get access to the feature and 50,000 do not (hypothetical). Only 30% of users with access ever turn the feature on.
  senior12 min

• Experimentation and causality

  An airline launches a lounge-subscription marketing campaign in one city (Izmir) and cannot randomise users. There are 20 comparable cities that did not get it.
  senior15 min

• Business judgment

  A marketing email campaign cost 81,000 and was sent to 450,000 customers, with 50,000 comparable customers held out at random (hypothetical figures).
  senior15 min

• DL, NLP and LLMs

  You must fine-tune a transformer classifier, but only 8 sequences of up to 512 tokens fit on one GPU at a time and the recipe you are reproducing used a batch size of 512.
  senior15 min

• DL, NLP and LLMs

  A team fine-tunes an instruction-following model on 20,000 labelled airline customer-service replies. Task accuracy on their evaluation set rises (hypothetical numbers) from 71% to 89%, and they want…
  senior15 min

• RAG and agents

  Your production RAG system uses a 1,024-dimension embedding model over 4 million chunks. The provider announces that the model will be retired in six months and offers a new model with a different…
  senior15 min

• RAG and agents

  An internal agent has 20 tools (booking lookup, fare rules, refunds, vouchers, email send, and so on), each with a short description and a JSON schema.
  senior15 min

• Production and system design

  A regulator asks your team to explain one specific loan decline from nine months ago: what exactly did the model see, which model version produced the score and why. Nobody can reproduce it.
  senior12 min

• Production and system design

  A customer asks to have their data deleted. You discover it appears in raw logs, a feature table, a training set used by three deployed models, a vector index for an assistant, and the prompts and…
  senior15 min

• Classical ML

  In a no-show model, the baggage-weight field is missing for 40% of high-risk-segment bookings and 5% of the rest (hypothetical). An analyst wants to drop rows with missing values;
  senior15 min

• Forecasting

  The airline launches a new route next quarter and the network team needs weekly demand forecasts and a first capacity plan, but the route has no history.
  senior15 min

• SQL and Python

  A daily table stores uniquecustomers per route per day. Someone builds weekly and monthly dashboards by summing the daily values, and the weekly numbers disagree with the numbers from the raw events.
  senior12 min

• Experimentation and causality

  A pricing-page test randomised 40,000 users (20,000 per arm). The analyst compared revenue per session across about 160,000 sessions with a two-sample test and got p = 0.03.
  senior12 minin review

• Experimentation and causality

  You randomise 50/50 on whether users are offered a new "trip planner" flow. Only 30% of users in the treatment arm actually open it; control users cannot access it.
  senior12 minin review

• Analytics

  A monthly cohort table (hypothetical, data through the end of April) shows customers by first-booking month and the share who booked again in month 1, 2 and 3 after joining.
  senior12 minin review

• Analytics

  Your customer dimension is a Type 2 slowly changing table customerscd(customerid, tier, validfrom, validto) with validto exclusive and 9999-12-31 for the current row.
  senior15 minin review

• Business judgment

  Leadership sets the goal "reduce route-level forecast MAPE by 20% this year" and ties bonuses to it. After six months the average MAPE across 200 routes (unweighted) dropped from 21.3% to 16.6%.
  senior15 min

• Communication

  The route-level weekly demand forecast has a MAPE of about 7% (hypothetical). An operations VP wants to use it for next-day staffing at flight level, where flights are around 150 passengers, and says…
  senior12 min

• Production and system design

  A ranking service must answer within 300 ms at the 99th percentile. For each request it calls five independent backend services in parallel (inventory, price, user profile, ML features, ad filters);
  senior15 min

• Statistics

  A small test on a new checkout button gives 12 purchases from 200 control visitors and 20 purchases from 210 treatment visitors (hypothetical).
  senior15 min

• Statistics

  An experiment's primary metric was flat. A stakeholder asks for a breakdown by 40 customer segments and finds four segments with p-values 0.004, 0.019, 0.038 and 0.047 (hypothetical;
  senior12 min

• Classical ML

  Marketing asks for "four customer segments" from a k-means model run on 120,000 loyalty members, using total spend (range 0 to 50,000), number of trips (0 to 30), days since last trip, and share of…
  senior15 min

• DL, NLP and LLMs

  You serve a 7-billion-parameter decoder-only model in 16-bit precision on one 80 GiB GPU for a support assistant with 8,192-token contexts. Throughput is far below the benchmark, and requests queue.
  senior15 min

• RAG and agents

  Whiteboard case. Design an internal RAG assistant for a company of 20,000 employees. Documents live in many systems (wiki, tickets, drive, HR) with their own permissions;
  senior35 min

• RAG and agents

  A research agent for an airline's network-planning team calls a search tool, a SQL tool and a spreadsheet tool in a loop, up to eight steps per task.
  senior15 min

• SQL and Python

  An analyst's daily-active-users query runs on events(userid, ts, eventname), a table of about 4 billion rows covering four years and laid out by date of ts (hypothetical sizes): It takes 40 minutes…
  senior15 min

• SQL and Python

  Whiteboard case. Design the nightly incremental pipeline that loads booking events from an operational system into an analytics table. Delivery is at-least-once, so duplicates occur;
  senior30 min

• Forecasting

  Your model's nominal 80% prediction intervals for weekly route demand cover the actual value only 62% of the time across 500 backtest forecast origins. Planners use the upper bound to set capacity.
  senior15 min

• Production and system design

  Whiteboard case. A payments-risk team and a recommendations team both want "fresh" per-user features, such as transactions in the last hour and average basket in the last 30 days, at serving time…
  senior35 min

• Business judgment

  A four-week pilot of a pricing-recommendation model on 20% of routes shows revenue +1.1% against control, with a 95% interval of -0.3% to +2.5%.
  senior15 min

• Analytics

  The airline's post-flight survey was sent to 50,000 passengers (hypothetical): 10,000 business-cabin and 40,000 economy. Response rates were 30% and 6%, so 3,000 business and 2,400 economy answered.
  senior15 min

• DL, NLP and LLMs

  You compare two prompts for an LLM that extracts the refund reason from customer messages, on the same 200 labelled messages. Prompt A gets 156 right (78%), prompt B gets 164 right (82%).
  senior15 min

• RAG and agents

  A support assistant answers policy questions with retrieval over the policy library. Policies are re-published under the same title, and the index contains the current version and up to two…
  senior15 min

• Experimentation and causality

  A new rebooking flow should cut support calls and must not increase flight cancellations through accidental self-service cancels. The cancellation rate in control is 2.0% (hypothetical).
  senior15 min

• Business judgment

  A retention programme calls customers ranked by an uplift model. A holdout experiment on 10,000 customers (hypothetical, 1,000 per decile of predicted uplift) shows the incremental number of…
  senior15 min