Practice
Workplace simulations and interview practice. Optional beyond the weekday lessons; studios embed selected scenarios. Everything simulated is labelled.
- 100scenarios
- 25capabilities covered
- 62 hof simulated work
- 150interview questions
Scenarios by capability100
Production incident
Senior code review
Stakeholder simulation
Messy real-world data
ADR centre
Live coding and SQL
- SQL gym: rolling and same-weekday features with window functions30 min
- Python gym: entity-safe lag and rolling features in pandas30 min
- SQL gym: consecutive nearly-full streaks and why the first query lies35 min
- Python gym: turning cumulative booking curves into pickup in pandas35 min
- SQL gym: booking pace against the same weekday last year40 min
ML system design
Ambiguous business case
Verbal defense
Statistical traps
Ablations
Decision theory and value
Robustness and stress tests
Paper reproduction
RAG evaluation workbench
LLM failure lab
- Unsupported citations and broken JSON in a claims assistant35 min
- Injected instructions and a looping helpdesk agent40 min
- A claims summariser that never saw its last document35 min
- A refund assistant that paid twice, in the wrong unit35 min
- An oversized voucher late in a long baggage-claim conversation40 min
Multi-agent company
End-to-end lifecycle
Golden test set and regression
Hiring committee
AI cost–quality
Agents vs workflows
Teach-back and Socratic interviewer
Executive presentation
Research reading protocol
- Read a churn paper before believing its abstract35 min
- Read an agent benchmark paper before adopting its framework35 min
- Does long context make retrieval obsolete? Read the claim40 min
- A language-model judge that agrees with humans 85% of the time35 min
- LoRA matches full fine-tuning: how much can 400 invoices show?35 min
Interview gym159
- A pricing test on one airline route returns p = 0.03 for the difference in conversion between two fare displays.mid5 min
- A bank's card-fraud rule catches 95% of fraudulent transactions and flags 2% of legitimate ones. Fraud is 0.5% of the 200,000 transactions processed per day, and the fraud operations team can review…mid6 min
- Using flightday(date, route, passengers, ...), write SQL for a 7-day rolling average of daily passengers per route. Then tell me what breaks if some dates are missing for a route.mid10 min
- For each route, return the three highest-revenue flight-days. Two flight-days on a route can have identical revenue.mid6 min
- Two queries in a code review. (1) A report meant to list every flight-day with its web promotion, if any, uses flightday f LEFT JOIN promotions p ON ...mid8 min
- In pandas you merge a bookings DataFrame (one row per booking) with a customer DataFrame on customerid. The result has more rows than bookings, and some bookings have no customer attributes.mid8 min
- A colleague computes average load factor per market in pandas with df.groupby("market")["loadfactor"].mean() and reports it as "market load factor". Some rows have a missing market.mid8 min
- Monday morning: the weekly bookings dashboard shows bookings down 8% week over week. Your director asks for an explanation by noon. How do you approach it?mid10 min
- On a route, load factor rose from 78% to 84% year over year, but revenue fell 3%. Capacity was unchanged. Decompose the change and tell me what you would show the route manager.mid8 min
- What is overfitting? How would you detect it? And in a production forecasting system, how would you tell overfitting apart from temporal leakage?mid10 min
- Your insurance-claims fraud classifier has 99.2% accuracy and 0.93 ROC-AUC. Fraud is 0.8% of claims. The investigation team can review 200 claims a week.mid8 min
- A teammate adds a target-encoded travelagencyid (about 4,000 agencies, many with fewer than ten bookings) to a no-show model. Validation ROC-AUC jumps from 0.71 to 0.86.mid10 min
- A colleague evaluated a daily passenger forecasting model with shuffled 5-fold cross-validation and reports 3% MAPE. What is wrong, and how would you evaluate it?mid8 min
- A retailer evaluates SKU-store daily demand forecasts with MAPE. Many SKU-days have zero sales. Which metric would you use instead, and why?mid8 min
- Your neural network's training loss stays flat from the first epoch, and on a second run it becomes NaN after a few hundred steps. Walk me through how you debug it.mid10 min
- Explain self-attention to an engineer who knows RNNs. Why did Transformers largely replace RNNs for language tasks, and what does attention cost?mid8 min
- You must route 50,000 customer emails a month into 12 categories for an airline's service team. You have 8,000 labelled emails. A colleague wants to fine-tune a large language model immediately.mid10 min
- Walk me through the components of a retrieval-augmented generation (RAG) system, and name one way each component typically fails.mid8 min
- Your vector search works well for natural-language questions but fails on queries like "fare basis YLOWTR" or "error code E-4012". Why, and what would you change?mid8 min
- Booking conversion on the website is 5%. You want to detect an absolute lift of 0.5 percentage points (to 5.5%) with a two-sided test at alpha 0.05 and 80% power.mid10 min
- A team ran an A/B test planned for 14 days but checked the p-value every day and stopped on day 6 when it dropped below 0.05. They want to ship.mid8 min
- Before launch, you replay 10,000 recent bookings through the new online no-show scoring service and through the offline batch pipeline used for training.mid10 min
- A manufacturing plant uses a model that outputs the probability a part is defective. Inspecting a part costs 20; shipping a defective part costs 500 (warranty and rework).mid8 min
- A marketing director expected a new email campaign to lift bookings. Your test result: +0.8% with a 95% confidence interval of -1.5% to +3.1%. How do you communicate this result?mid8 min
- Tell me about a time an analysis or model you delivered turned out to be wrong. What happened and what did you change? (Behavioural: answer using STAR.)mid8 min
- An airline's baggage team reports (hypothetical figures) 18 mishandled-bag claims on 100,000 bags last quarter and 31 claims on 120,000 bags this quarter.mid8 min
- A dashboard shows the cancellation rate by fare class from a table bookings(bookingid, fareclass, status).mid8 min
- On 30 June an analyst pulls a cohort table of 90-day repeat-booking rate by month of first booking and shows the following hypothetical figures: Jan 17.1%, Feb 16.8%, Mar 17.0%, Apr 15.2%, May 11.1%,…mid10 min
- A model predicts whether a customer will rebook within 90 days. The table has 10,000 rows from 2,000 customers (five bookings each, on average).mid8 min
- A new daily passenger-demand model has an MAE of 6.0 passengers on the latest validation week, and the dashboard says its error is "71% lower than the naive forecast" (the naive forecast being…mid8 min
- The VP of customer experience messages you: "Why are customers so unhappy about baggage? I need this for Thursday's leadership meeting." You have survey data, claims data and a customer-service…mid8 min
- An LLM step that extracts a fare class from customer emails returns different answers when the same email is sent twice. A teammate says "just set temperature to 0 and it is deterministic".mid8 min
- You are chunking 10,000-token policy documents for a RAG system. A teammate proposes 500-token chunks with a 100-token overlap.mid8 min
- You model ancillary revenue per booking with linear regression. Revenue is strongly right-skewed, so you fit it on the log scale and get a good fit.mid8 min
- A weekly demand forecast for a route has been reviewed with MAE and it looks acceptable. A planner says "it keeps coming in low".mid8 min
- A weekly count of baggage-delay complaints averages 40 with a variance of 160 across weeks. A colleague builds a control chart with limits at the mean plus or minus three Poisson standard deviations…mid8 min
- Bookings are stored with UTC timestamps. The daily bookings report for the London office groups by the UTC date, and the commercial team says Sunday numbers do not match their own system, especially…mid8 min
- Your company's experimentation platform has just been rebuilt. Before trusting it, you are asked to run 100 A/A tests (both arms see the same experience) at a 5% significance level.mid8 min
- You have events(userid, eventname, eventts) with names 'search', 'select' and 'pay'. Write SQL for the funnel search to select to pay: users reaching each step, step conversion, in the correct order.mid10 min
- A team proposes a forecasting project. Hypothetical estimates: build cost 250k, run cost 90k per year, benefit of 400k per year if it works, with a ramp to half the benefit in year 1, and a 60%…mid10 min
- A colleague's slide shows on-time performance for last quarter and this quarter (96% and 98%) as a bar chart with the vertical axis starting at 94%. The title is "On-time performance doubled".mid8 min
- Your weekly batch job scores 2 million customers and appends rows to a scores table consumed by the CRM.mid12 min
- A league table ranks two regional carriers on on-time performance (OTP), defined as the share of operated flights that arrive within 15 minutes of schedule (hypothetical figures).mid10 min
- You fit a logistic regression for flight cancellations on bootstrap resamples of the same data. Two features, distancekm and scheduledminutes, correlate at 0.98.mid10 min
- You must classify baggage-damage photos into five damage types with 1,200 labelled images. The photos come from about 400 claims, three photos per claim.mid12 min
- For one test question in a RAG system, the retriever returns ten chunks. The chunks that a labeller marked relevant sit at ranks 2, 5 and 9;mid10 min
- Given bookings(customerid, bookedon) with several rows per customer, write a monthly cohort retention table: cohort = month of a customer's first booking, columns = months since the first booking,…mid12 min
- You build one model to forecast daily passengers 1 to 14 days ahead. A teammate's feature set contains passengerslag7, passengerslag14, holidayflag, competitorfaretoday and weatheractual.mid10 minin review
- A monitoring job reports a Population Stability Index of 0.066 for the feature daystodeparture and 0.317 for bookingchannelshare. The model's score distribution looks unchanged.mid10 min
- A satisfaction analysis regresses a passenger's 1-10 satisfaction score on the delay they typed into a post-flight survey ("how many minutes late was your flight?").mid10 min
- A teammate classifies 100 labelled claims as fraudulent or not using 2,000 candidate features extracted from claim text and metadata.mid10 min
- A marketing analyst finds that the weekly "cheap flights" search-interest index and weekly bookings, over 104 weeks, have a correlation of 0.95 (both series are simulated for this exercise).mid10 min