Interview guides · Blackstraw.ai

Blackstraw.ai Data Scientist Interview Questions & Guide 2026

Blackstraw.ai data scientist prep on data pitfalls: selection bias, target leakage, and checking an outlier's provenance before you drop it.

The 60-second brief

As a Data Scientist at Blackstraw.ai, you are at the intersection of advanced research and practical engineering. You will be responsible for designing and operationalizing AI systems that solve complex, real-world business challenges for global enterprises. Your work is not just about building models in isolation; it is about creating scalable, end-to-end solutions that drive measurable impact in domains like Predictive Analytics, Computer Vision, and Generative AI.

The boundary with engineering varies enough to be worth asking about directly. Some seats end at the analysis and the writeup; others carry feature pipelines, scheduling and alerting, which decides whether production data plumbing belongs in your prep at all.

Focus areas

Separate contracted seats from actively used seats
Measure churn only on renewal-eligible accounts
Power experiments for heavy-tailed account revenue

Overview

This role is inherently cross-functional and fast-paced. You will collaborate with engineering and product teams to translate raw data into actionable business stories, bridging the gap between technical complexity and stakeholder value. Success here requires a blend of deep domain expertise and a "learn-by-doing" mindset, as you will often be involved in the entire lifecycle of a project—from conceptualization and statistical modeling to production deployment.

The Blackstraw.ai interview process

Blackstraw.ai candidates report 6 rounds · ≈ 4-6 weeks. The stages below are what candidates describe, not a published process.

Initial Screening

Whoever runs this call is usually not a practitioner. They take notes, and a hiring manager skims those notes later, so the real question is whether your work survives being written down by someone outside the field. Test every project sentence against that: could a non-specialist repeat it correctly without knowing what a propensity score is? Carry a plain-language version of each project and one reason you want this particular role that you could not copy onto another application. Vagueness at this stage reads as inexperience, even when the underlying work was genuinely deep.

What to demonstrate

Whether a non-specialist can restate your projects accurately, since their paraphrase is what reaches the hiring manager
Whether your reason for wanting the role points at the work itself rather than the company's reputation
Whether your language signals the level being screened for: what you decided yourself versus what you were handed

How to prepare

Write a two-sentence, jargon-free version of each major project: the question nobody could answer, and the decision your work changed. Read it to someone outside data and have them repeat it back
Point your 'why this role' answer at something concrete in the job description or the product surface you would be working on, and keep it to two sentences
Have two questions ready about measurement: which metric the team is held to, and who acts on an analysis once it lands

Technical Discussions

Before anything else, this round is a reading test. You are given a small schema and a question phrased in business language, and most of the difficulty sits in the gap between them. Who counts as an active user, does a refunded order still count as an order, is that date column an event time or a load time. Weak answers start typing immediately and compute something precise about the wrong population. Strong ones pin the definition in one sentence, name the column that encodes it, then write the query. On a timed assessment with nobody to tell, write the definition in a comment anyway.

What to demonstrate

Whether an ambiguous term becomes a specific column and filter before any computation happens
Whether you read the schema for keys and cardinality rather than only for column names
Whether the result answers the question at the grain it was asked at, per user or per session or per day

How to prepare

Take three metrics you already use and write down the exact filter and exact grain behind each, then practise stating one of them in a single sentence out loud
On a schema you have never seen, spend the first minute writing what one row of each table means and which key it is unique on, then predict which joins can duplicate rows
Rehearse a version where the definition changes halfway through, and edit the query you have instead of starting over

Project-Based Discussions

An extra round usually exists because something is still open after the standard loop: a skill the earlier interviews did not sample, a level decision, or two interviewers who disagreed. It is rarely a rerun of what you already did well. Ask the recruiter who you are meeting, what function they sit in, and how long the session runs. That is an ordinary scheduling question, and the answer changes what you should prepare. What separates a strong candidate here is treating the round as a fresh evaluation with its own bar, rather than assuming earlier performance carries you through or sinks you.

What to demonstrate

Whether you can answer well on ground the earlier rounds did not cover, without leaning on what you already said to someone else
Consistency of the facts in your stories: the same sample size, timeframe, team size and scope of your own role as in earlier conversations
How you handle an unfamiliar format live, including whether you ask what kind of answer is wanted before producing one

How to prepare

Ask the recruiter for the interviewer's function, the length, and whether to expect a coding surface, a discussion, or a presentation. Preparing for a 30 minute conversation with a partner team is not the same work as preparing for a 60 minute technical block.
Write out what each earlier round actually covered, then list the two or three areas nobody probed. That gap is the most likely subject of the extra round.
Re-read the numbers in the project stories you have already told, so a second telling does not quietly contradict the first.

System Design Conversations

Because the format is not fixed, prepare the reasoning rather than the ritual. Nearly every version of this round draws on the same underlying material: a design you can defend, a metric you can define exactly, an analysis whose assumptions you can state out loud. Only the wrapper changes, whether that is a take-home, a live case, a deep dive on past work, or a rough estimate on a whiteboard. Answers rehearsed to fit one shape stall the moment the shape differs. Practise naming the assumption behind a number, then saying how much the conclusion moves if that assumption is wrong.

What to demonstrate

Whether your justification for a method survives the question 'why not the simpler thing', including when the simpler thing would have worked
Precision under pressure: what exactly counts as an active user, a conversion or a success, over what window, with what exclusions
Whether you carry an argument through to a recommendation instead of stopping at a list of tradeoffs

How to prepare

For each project you plan to mention, write the metric definition in one sentence: numerator, denominator, time window, exclusions. Say it out loud once, because vagueness shows up in speech before it shows up on paper.
Rehearse the same project at three lengths: two minutes, ten minutes, and a deep dive on one technical decision. Cutting live is harder than it sounds.
For your headline result, write down what would have had to be true for it to be wrong, and how you ruled that out.

Granular Technical Deep Dives

A handful of shapes account for most of what gets asked in this format: a ranking or deduplication inside groups, a running or rolling total, a period-over-period comparison, and a cohort tracked forward over time. Recognising the shape quickly is most of the speed here; deriving it from scratch while a clock runs is where the time goes. Know that a window function keeps every row while a GROUP BY collapses them, and know which one the question needs. If the exercise is in Python instead of SQL, the same shapes arrive as groupby with transform, shift and merge, and the same grain mistakes are available.

What to demonstrate

Whether you reach the right construct without a detour, such as ROW_NUMBER over a partition to deduplicate instead of a self-join against a MAX subquery
Whether you know what your window frame actually is, since adding ORDER BY inside OVER changes the default frame and silently changes a running total
Whether the thing runs. A near-miss that throws an error scores below a plainer query that returns the right rows.

How to prepare

Write each of the four shapes once from memory against a small schema and keep the working version somewhere you will reread it: dedupe with ROW_NUMBER, a running total, a month-over-month change with LAG, and a retention table
Compute one running total twice on data with tied timestamps, once on the default frame and once with ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW, and look at where the two disagree
If Python is on the table, rebuild the dedupe and the running total with groupby and cumsum, then assert the two implementations return identical rows

Final Leadership Discussions

Where a loop ends with a senior leader, that conversation is rarely another skills test. The technical signal already exists by then, so the questions tend to open up: what you would look at first, where a metric you have heard about could mislead, what you would push back on. The decision being made is scope, which in practice means level and how much you would be trusted to own unsupervised. Treating it as a formality is the usual mistake. An open question late in the day is still being scored, and a vague answer reads as someone who has not run anything themselves.

What to demonstrate

Whether your view of the business has anything specific behind it, given that you are working only from what is public and are expected to say so
Whether the scope of work you describe owning matches the scope of the role, instead of sitting a level below it
Whether you can disagree with something concrete and stay useful about it, rather than agreeing with everything said in the room
Whether your questions are ones only this person could answer, as opposed to ones the recruiter already covered

How to prepare

Build one view you could defend for two minutes using only public information: what the funnel probably looks like, which metric likely drives decisions, and where that metric could mislead. Being wrong for a stated reason survives this round; having no view does not
Write down the largest piece of work you have owned from question to decision, who else touched it, and what you decided alone, then check that it reads at the level you are interviewing for
Prepare one thing you would want changed if you joined and phrase it as a question rather than a verdict, so it opens a conversation instead of closing one

Where candidates lose points

Comparing accounts that received a sales or customer-success touch against those that did not

Assignment of coverage is deliberate and pulls in both directions at once: the largest accounts get a named owner because they are valuable, and the accounts showing distress get one because they are at risk. The comparison therefore mixes a strong positive selection with a strong negative one, and the naive estimate can come out with either sign depending on which assignment rule dominated during the period examined. Nothing about matching on observed size fixes this, because the risk signal that triggered coverage is usually the same signal that predicts the outcome. It needs either an actual randomised or staggered rollout of coverage, or a design built on a capacity constraint or territory boundary that assigns coverage for reasons unrelated to account health.

Reading consumption metrics before the metering lag window has closed

Usage pipelines land late and correct themselves, which is exactly what is_restated and restated_at record. A dashboard queried on day T sees a partially populated tail for the last several days, so the most recent points always slope downward and always look like a regression. Analysts then explain the artefact, and sometimes ship a change to fix it. Establish the empirical settling time by measuring how much a given usage_date's total moves between first_written_at and its final value, exclude that many trailing days from every reportable figure, and never compare a fresh period against a settled one.

Defining the cohort on a post-treatment condition

Ask how rows entered the table. Filtering on something that treatment itself influences, such as users who finished onboarding or accounts still active at ninety days, breaks comparability between arms; define the population at an entry point that precedes exposure and keep everyone in it.

Over-explaining the method and under-explaining the implication

Lead with the answer and what you would do about it, then give the approach when asked. Roughly one sentence of method per three of implication is the right ratio for a stakeholder-facing answer; the interviewer already knows what a regression is.

Questions & practice

How did you deploy your model into a production environment?

Medium · machine learning and modelling

Approach

Pick an evaluation metric that matches the cost of each error type, not a default.
Check what information would not exist at prediction time, and exclude it.
Say how the offline result would be validated online before it is trusted.

Follow-up

What would you monitor after launch to know the model is still valid?
Where could label leakage enter this setup?

Describe the mathematical intuition behind a specific algorithm you have used in a project.

Medium · machine learning and modelling

Approach

Check what information would not exist at prediction time, and exclude it.
Frame the prediction: the label, the moment of prediction, and the action it triggers.
Pick an evaluation metric that matches the cost of each error type, not a default.

Follow-up

How would you choose the decision threshold, and who owns that choice?
What would you monitor after launch to know the model is still valid?

How do you measure the success of a model in a business context?

Medium · machine learning and modelling

Approach

Frame the prediction: the label, the moment of prediction, and the action it triggers.
Set a baseline first, so any model has something honest to beat.
Pick an evaluation metric that matches the cost of each error type, not a default.

Follow-up

Where could label leakage enter this setup?
What would you monitor after launch to know the model is still valid?

Bootstrap a confidence interval for net revenue retention

Medium · bootstrap · retention · resampling

You have one row per account with arr_start_cents (ARR twelve months ago) and arr_end_cents (ARR today, zero if churned), covering the fixed cohort of accounts that had ARR twelve months ago. Net revenue retention is sum(arr_end_cents) / sum(arr_start_cents). Write a nonparametric bootstrap from scratch, without scipy.stats.bootstrap: resample accounts with replacement, recompute the ratio of sums on each resample, and return the point estimate with a 95% percentile interval from 10,000 resamples. Also report the interval you would get from the mean of per-account ratios, and explain the difference.

Approach

Resample the account, because the account is the unit the estimand is defined over. One bootstrap draw is a vector of account indices and both numerator and denominator are recomputed from that same draw; resampling the two sides independently destroys the within-account correlation that makes a ratio estimator stable.
Vectorise the draws: idx = rng.integers(0, n, size=(B, n)), then end[idx].sum(axis=1) / start[idx].sum(axis=1). A 10,000 by n index matrix is usually far cheaper than a Python loop; if the matrix is too large for memory, chunk over B rather than reverting to a loop.
Take the interval from np.quantile(ratios, [0.025, 0.975]). The percentile interval differs from estimate +/- 1.96 * bootstrap SE whenever the resample distribution is skewed, which it will be here, and the skew is the thing you want represented.
Compute the mean-of-ratios version on the same resamples, and state the exact relationship rather than guessing which of the two is larger. With r_i = arr_end_i / arr_start_i, the ratio of sums is the arr_start-weighted mean of exactly those r_i, so sum(end)/sum(start) - mean(r) = Cov(arr_start, r) / mean(arr_start) using the population covariance. The gap is positive when larger accounts retain and expand better than smaller ones, and negative when they do not; a cohort whose small accounts churn at a higher rate has positive covariance, which puts the mean of per-account ratios BELOW the ratio of sums. Requires arr_start_i > 0 for every account, which the fixed-cohort definition guarantees; r_i is floored at 0 and unbounded above, so a handful of 4x expansions among small accounts can flip the sign. Compute the covariance and report it instead of asserting a direction.
Report the interval width beside the concentration of the cohort. If the largest account is 12% of starting ARR, a narrow interval is evidence that the resampling unit is wrong rather than evidence that the estimate is precise.

Worked solution (30 min)

start = df.arr_start_cents.to_numpy(float); end = df.arr_end_cents.to_numpy(float); n = len(start); point = end.sum() / start.sum()
rng = np.random.default_rng(7); idx = rng.integers(0, n, size=(10_000, n)); ratios = end[idx].sum(1) / start[idx].sum(1)
lo, hi = np.quantile(ratios, [0.025, 0.975]); return point, lo, hi
per_acct = end / start; mean_point = per_acct.mean(); mean_boot = per_acct[idx].mean(1); compare np.quantile(mean_boot, [0.025, 0.975]) against (lo, hi), and report np.cov(start, per_acct, ddof=0)[0,1] / start.mean() as the quantity that accounts for the gap between the two centres.

Expected result: A point estimate equal to sum(end)/sum(start) on the unresampled data — the bootstrap mean is close but must not be substituted for it — and a percentile interval that is asymmetric around the point when expansion is skewed. The mean-of-ratios interval is centred on a different estimand, and which side it falls on is decided by Cov(arr_start, r): it sits below the ratio of sums in the common case where small accounts churn at a higher rate than large ones, and above it when the large accounts are the ones shrinking. Its width is governed by the spread of per-account growth, while the ratio-of-sums width is governed by how concentrated starting ARR is, so neither interval is reliably the wider one.

Verify it yourself

point equals sum(end)/sum(start) computed directly on the original arrays, and point - per_acct.mean() equals np.cov(start, per_acct, ddof=0)[0,1] / start.mean() to floating-point tolerance. That is an algebraic identity, not an empirical claim: a mismatch means the per-account ratios or the weights are wrong, and the sign of it tells you which of the two numbers is larger before you look.
np.median(ratios) sits within a small fraction of a point of the point estimate; a large gap indicates a wrong resampling unit or a transposed index matrix.
Raising B from 10,000 to 20,000 moves the endpoints by far less than the interval width.
Setting every arr_end equal to its arr_start gives point == 1.0, per_acct.mean() == 1.0, a zero covariance term and a zero-width interval.

Follow-up

The cohort has 800 accounts and the largest is 12% of starting ARR. How much do you trust a percentile interval here?
How would you extend this to an interval on the year-over-year change in NRR?
Two accounts merged mid-window and one contract was co-termed into the other. How do you keep the cohort fixed?

How do you select the right algorithm for a specific predictive modeling problem?

Medium · sql and data manipulation

Approach

Say which table is the grain you start from, and join outward from it.
Compute rates by summing numerator and denominator separately, never by averaging rates.
Handle the rows that do not match: a LEFT JOIN with a NULL check is usually the question.

Follow-up

How does the query change if the join becomes one-to-many?
What breaks if events arrive late or out of order?

Running commitment burn-down and the date consumption crosses it

Medium · window functions · running totals · committed consumption

fct_subscription_period gives committed_amount_cents, term_start_date and term_end_date for each account's current version where pricing_model = 'committed_consumption'. fct_usage_daily gives account_id, workspace_id, sku_code, usage_date and net_amount_cents, at one row per workspace and SKU per day. Inside each account's current term, return the running total of net_amount_cents by usage_date, the first usage_date on which that running total reaches committed_amount_cents, and the fraction of the term elapsed at that point. Accounts that have not reached their commitment must still appear, with a null crossing date.

Approach

Collapse usage to one row per (account_id, usage_date) first. The fact is grained by workspace and SKU, so a raw running total leaves several rows per date and the first-crossing date becomes dependent on the arbitrary order of rows inside that day.
Restrict to the term with usage_date BETWEEN term_start_date AND term_end_date on the account's current row, so no prior term's consumption leaks into this term's burn-down.
Compute SUM(net_cents) OVER (PARTITION BY account_id ORDER BY usage_date ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW). Name ROWS explicitly: the default frame is RANGE, which includes all peer rows at the same ORDER BY value and would hand back the whole day's total on each of that day's rows.
Extract the crossing with MIN(usage_date) FILTER (WHERE running_cents >= committed_amount_cents) grouped per account, which returns NULL for accounts still under commitment instead of dropping them.
Elapsed fraction is (crossing_date - term_start_date)::numeric / NULLIF(term_end_date - term_start_date, 0); Postgres date subtraction yields whole days, so the guard matters for same-day terms.
Exclude the trailing days still inside the metering settling window, measured from first_written_at against restated_at, and say how many days you cut and why.

Worked solution (30 min)

Build terms: SELECT account_id, committed_amount_cents, term_start_date, term_end_date FROM fct_subscription_period WHERE is_current AND pricing_model = 'committed_consumption' AND committed_amount_cents IS NOT NULL.
Build daily: join fct_usage_daily to terms on account_id with usage_date inside the term, then GROUP BY account_id, usage_date summing net_amount_cents.
Add running_cents with SUM(...) OVER (PARTITION BY account_id ORDER BY usage_date ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW).
Aggregate per account with MIN(usage_date) FILTER (WHERE running_cents >= committed_amount_cents) AS crossing_date, then compute the elapsed fraction with the NULLIF guard.
LEFT JOIN that summary back onto terms so every committed account appears, and verify the last running_cents per account equals a plain SUM over the term.

Expected result: A per-account-per-day series of running_cents inside the current term, plus one summary row per committed account carrying crossing_date and term_elapsed_fraction, with crossing_date NULL exactly for accounts whose final running total is below committed_amount_cents.

Verify it yourself

The maximum running_cents per account equals SUM(net_amount_cents) over that account's term computed independently.
crossing_date is NULL if and only if the final running total is below the commitment.
term_elapsed_fraction falls in [0, 1] for every non-null crossing date.
running_cents is non-decreasing within each account, which fails immediately if a negative credit row was included without comment.

Follow-up

Turn this into an end-of-term overage forecast. What breaks if you extrapolate a linear run rate on a consumption product?
An account crosses its commitment at 40 percent of the term. Is that an expansion signal or a billing-surprise risk, and what would you check to tell them apart?

How do you approach feature extraction in text data?

Medium

Approach

Say what you would check first and why it is the highest-information step.
State your assumptions explicitly before working the problem.
Work from the decision backwards to the evidence you would need.

Follow-up

How would you know your answer was wrong?
What assumption would you test first?

Walk us through a recent data science project you led from end to end.

Medium

Approach

Work from the decision backwards to the evidence you would need.
State your assumptions explicitly before working the problem.
Say what you would check first and why it is the highest-information step.

Follow-up

What assumption would you test first?
How would you know your answer was wrong?

Triage a sample ratio mismatch before reading the result

Easy · srm · assignment · instrumentation

An account-randomised test assigned arms 50/50 by hashing account_id. Exposure logging recorded 4,200 accounts: 1,987 in treatment and 2,213 in control. A chi-square goodness-of-fit test against the intended 2,100/2,100 split gives chi-square(1) = 12.2, p about 0.0005. The analyst reports the primary metric up 6% in treatment and asks to ship. Explain what the imbalance implies about that 6%, rank the causes you would investigate first given this domain's data, and name for each the single query against fct_api_request or dim_account that confirms or eliminates it.

Approach

State the rule before diagnosing anything: a sample ratio mismatch below a 0.001 alarm threshold invalidates the readout. Whatever mechanism removed 113 accounts from one arm almost certainly removed them non-randomly, which makes it a selection effect on the outcome. The 6% is not to be adjusted, caveated or shipped; it is unusable until the mechanism is named.
Verify the test is testing the right thing. chi-square = (1987 - 2100)^2 / 2100 + (2213 - 2100)^2 / 2100 = 6.08 + 6.08 = 12.16 on one degree of freedom. Confirm the denominator is the intended assignment count rather than the observed total, and that you are testing assignment rather than an analysis population that has already been filtered.
Rank causes by how this domain actually breaks rather than by textbook order: is_internal accounts filtered after assignment instead of before; exposure logged on a code path that returns 5xx more often in one arm, silently dropping those accounts; assignment recorded at first request but exposure at a later event, so accounts that churned in between appear in only one arm; dim_account type-2 versioning giving one account_id two surrogate keys and two hash inputs; deployment_model = 'self_hosted' accounts that never emit exposure telemetry at all.
Attach one discriminating query to each: count arms before the is_internal filter; compare the http_status >= 500 rate by arm on the exposure endpoint in fct_api_request; compare the distribution of created_at and churned_at by arm in dim_account; count distinct account_sk per account_id inside the experiment window; cross-tabulate arm against deployment_model.
Compare counts at three fixed stages, assignment, first exposure, and analysis population, and localise the divergence to one of them. The stage where the arms first separate names the subsystem, and everything downstream of it is a symptom.
Report the outcome as abort, fix, rerun, and be explicit about what survives: the variance estimate for re-powering, the instrumentation fix, and nothing whatsoever about the effect size.

Worked solution (20 min)

Recompute the statistic: two terms of 113^2 / 2100 = 6.08, total 12.16, p about 0.0005 on one degree of freedom, below the 0.001 alarm line.
Pull arm counts at assignment, at first exposure, and in the analysis population, and find the first stage at which they diverge.
At that stage, cross-tabulate arm against is_internal, deployment_model, and the 5xx rate on the exposure endpoint.
Write the abort note naming the mechanism, the corrected code path, and the rerun date, with the effect estimate explicitly withheld.

Expected result: chi-square(1) = 12.16, p about 0.0005, so a 0.001 alarm threshold fires and the readout is abandoned. The divergence is localised to a single pipeline stage by comparing arm counts at assignment, at exposure and at analysis, and the most likely causes in this domain are post-assignment internal-account filtering or exposure logging that fails asymmetrically on 5xx responses.

Verify it yourself

A 113-account shortfall on 2,100 expected is a 5.4% deficit in one arm, which under a sound hash is roughly a 1-in-2,000 event. That is what the 0.001 threshold is calibrated to catch.
Counts at the assignment stage should already be balanced if the bucketing is correct. If they are not, the bug is in the hash or the salt, not in the logging.
Any fix that restores balance by filtering on a variable measured after exposure has conditioned on a post-treatment variable and reintroduces the bias it appears to remove.

Follow-up

The imbalance disappears once you restrict to accounts with at least one successful request. Does that fix the experiment or confirm the bug?
What alarm threshold would you set for this check, and why is 0.05 the wrong one for a diagnostic you run on every experiment every day?
How would you detect a mismatch confined to one segment while the overall split looks clean?

Billable units per account jumped while nothing shipped

Medium · retries · idempotency · guardrail metrics

Billable units per paying account rose 22 percent month over month with no pricing or packaging change. From fct_api_request (account_id, endpoint, http_status, is_retry, idempotency_key, traffic_class, billable_units, request_at) and fct_usage_daily (account_id, sku_code, usage_date, billable_quantity), determine how much of the rise is delivered value and how much is duplicated work. Deliverable: the decomposed figure, the accounts it concentrates in, and a recommendation on whether to report the 22 percent at all.

Approach

Check the guardrail before the headline: compute the customer-visible server error rate per account for both months, numerator http_status >= 500 and denominator excluding synthetic_monitor and load_test. Metered volume rising alongside an error rate is the known failure mode in this domain.
Deduplicate logical work by counting billable_units once per (account_id, idempotency_key) at the first successful request rather than once per row. Rows with a null idempotency_key cannot be deduplicated, so report their share as an explicit uncertainty band instead of assuming they are all unique.
Split by traffic_class before interpreting anything, because one change to a continuous-integration configuration can multiply request volume overnight without a human deciding anything about the product.
Test concentration: compute the per-account distribution of the increase and its top-decile share. A rise carried by a few accounts scaling one batch job is a different finding from a broad shift and gets a different recommendation.
Reconcile against fct_usage_daily for the same accounts and dates, and be ready to explain the expected gap in two sentences: request rows include retries and failures carrying zero billable_units, and the usage table restates after first write.

Follow-up

A client retrying a request the server already completed produces duplicate billed work. What protocol or product change removes that, and what would you measure to confirm it worked?
If the duplicated volume was genuinely invoiced, what should finance do, and how does that change what belongs in the metric?

Suggested practice checklist

For a candidate whose interviews will centre on A/B testing, metric movement and causal claims. Design comes before arithmetic, arithmetic before analysis, and the week ends by rehearsing the readout rather than the derivation.

Day 1 — Design one test end to end on paper

Take a single feature change and write the full design: randomization unit, the exact point of exposure, the primary metric with its grain, guardrails, allocation, planned duration, and the decision rule committed before any data exists.
Write why the randomization unit must sit at or above the level where treatment can spill over, and give one case where user-level randomization is still contaminated (shared accounts or devices, or two participants in the same marketplace).
State in advance what you will do if the primary metric is flat while a secondary metric is significant.

Deliverable: A one-page test design with a decision rule written before launch.

Day 2 — Power arithmetic until it is automatic

Compute required sample size per arm for a binary metric with the normal approximation, n is approximately 2 times (z for alpha/2 plus z for power) squared times p(1 minus p) divided by delta squared, for baselines of 2, 10 and 40 percent at a 5 percent relative lift, and note that for a fixed relative lift the requirement falls as the baseline rises because delta grows proportionally with p.
Redo the calculation for a continuous metric using variance in place of p(1 minus p), and show why a heavy-tailed quantity such as revenue per user needs either far more traffic or a capped version with a stated cap.
Convert one of the results into weeks given a weekly eligible traffic figure, then list the two honest ways to shorten it (accept a larger detectable effect, or reduce variance) and write why quietly lowering the power target is a decision to miss more real wins, not a speedup.

Deliverable: A small script or sheet that maps baseline, minimum detectable effect, alpha and power to sample size and weeks, cross-checked against a published calculator.

Day 3 — Variance and the unit-of-analysis problem

Take a ratio metric whose denominator is not the randomization unit (clicks per session, randomized by user) and compute the standard error twice, once naively at session level and once by the delta method or a user-level bootstrap, then record how much the naive version understates it.
Implement CUPED on simulated data: choose a pre-period covariate X measured before assignment, estimate theta as Cov(Y, X) divided by Var(X), and analyse Y minus theta times (X minus its mean) in place of Y. Confirm the variance of the adjusted outcome equals the raw variance multiplied by one minus the squared correlation between Y and X, so a correlation of 0.45 removes about 20 percent of the variance and not 80.
Now run that simulation a few hundred times and confirm the adjusted effect estimate is unbiased for the same effect rather than numerically identical to the raw one. Within any single run the two differ, sometimes by a large fraction of the true effect, because the two arms' pre-period covariate means never coincide exactly in a finite sample; they agree in expectation, which is the property that matters and the one to state out loud.

Deliverable: A notebook showing the adjusted estimator with a measurably smaller variance than the raw one, plus a repeated-simulation table showing the two estimators agreeing on average while differing run by run.

Day 4 — Validity threats you can actually test for

Run a sample ratio mismatch check as a chi-square goodness-of-fit test against the intended allocation, and write the three causes you would chase first (assignment logged before exposure, an arm-specific redirect or load failure, bot filtering applied asymmetrically).
Simulate peeking: generate A/A data, test daily at alpha 0.05 across 14 looks, record the inflated false positive rate, then apply an alpha-spending boundary or commit to a fixed horizon and confirm the rate returns to nominal.
Write how you would separate a novelty effect from a durable lift using the treatment effect plotted against days since first exposure, and what shape would change your recommendation.

Deliverable: One table showing the peeking false positive rate before and after correction, plus a written SRM triage list.

Day 5 — When randomization is not available

Write the identifying assumption for difference-in-differences (parallel trends in the absence of treatment), then plot pre-period trends for two candidate control groups and justify rejecting one of them.
Design a switchback test for a change where user-level randomization would leak across participants, choosing a time-block length against the carryover you expect and saying how you would detect carryover in the data.
List what an interrupted time series or a synthetic control buys you and the one thing neither can rule out: an unobserved shock that coincides with the launch.

Deliverable: A one-page memo recommending a single quasi-experimental design and naming its weakest assumption explicitly.

Day 6 — The readout query

Write the assignment-to-exposure join that returns exactly one row per unit per experiment, and handle units appearing in both arms by excluding and counting them rather than silently keeping one.
Compute the per-arm metric, its variance and the relative lift with a confidence interval in SQL, then reproduce the identical numbers in a notebook as a cross-check.
Add a segment breakdown and write the sentence that keeps it from being p-hacking: segments declared in advance, everything else reported as exploratory and corrected for multiplicity.

Deliverable: A single query that outputs the full readout table, matched to a notebook recomputation.

Day 7 — Present it to someone who will not read the appendix

Give a 10-minute readout of a real or simulated experiment in the order decision, number, uncertainty, caveat.
Have your listener ask "can we ship it" in the case where the primary is flat and a guardrail moved, and answer with a recommendation rather than a request for more data.
Rewrite your opening line so the recommendation lands before any methodology.

Deliverable: A one-page readout whose first line is the recommendation.

Ownership when the result is inconvenient

Sometimes the honest read is that the initiative did not work, and the person who commissioned the analysis was hoping otherwise. Interviewers want to know whether you softened it. Prepare the case where you delivered an unwelcome result, how you presented the uncertainty without hiding behind it, and what the team did next.

How do you handle concept extraction in large, unstructured datasets?

Medium · behavioural and stakeholder questions

Approach

Quantify the outcome, including what you would not claim credit for.
Pick a story where you drove the decision, not one where you observed it.
Close with what you would do differently, concretely.

Follow-up

What would you do differently if you ran that project again?
How did you know the outcome was caused by your change?

How do you handle overfitting in deep learning models?

Medium · behavioural and stakeholder questions

Approach

Quantify the outcome, including what you would not claim credit for.
Pick a story where you drove the decision, not one where you observed it.
State the situation in two sentences and spend the rest on your reasoning.

Follow-up

What did you decide not to do, and why?
What would you do differently if you ran that project again?

State the measured impact of your own work honestly

Hard · impact · causal inference · self-assessment

You are asked for the business impact of a renewal-risk worklist you shipped nine months ago. Customer success used it, and renewals in the covered segment came in 6 points above the prior year. Coverage was assigned by the team itself: they worked the top of your list and also the accounts they were already worried about. Produce the impact claim you are willing to defend, the number you refuse to claim, and the design you would have asked for at the start.

Approach

The interviewer is probing whether you can separate what you shipped from what you caused, and whether you would have built the measurement in rather than reconstructing it afterwards. Both halves are being scored.
Name the confound precisely. Coverage assignment is doubly selected: the largest accounts get an owner because they are valuable, and distressed accounts get one because they are at risk. The naive covered versus uncovered comparison mixes a strong positive selection with a strong negative one and can come out with either sign depending on which rule dominated. Matching on account size does not fix it, because the risk signal that triggered coverage is the same signal that predicts the outcome.
Split the claims by what each needs to be true. Ranking quality is defensible from precision at k on out-of-time renewals. Adoption is defensible from timestamps showing what share of listed accounts were contacted. The outcome claim is not defensible without a design, and saying so is the point of the exercise.
Look for identification before giving up on it. A capacity cut-off, a territory boundary, or a period in which the list existed but was unstaffed can assign coverage for reasons unrelated to account health, and any of those supports a bounded estimate.
State the design you would ask for now and its price: a randomly withheld slice of the list, held for two renewal quarters, with the expected cost in renewals stated openly. That cost is what it takes to be able to answer this question at all.
Give a bounded number rather than none. Six points with an explicit statement of how much of it you can attribute is more useful than either claiming the whole figure or declining to quantify anything.

Follow-up

Your manager wants the 6 points in a promotion packet. What wording do you accept, and what do you strike?
What would have had to be true for the naive covered versus uncovered comparison to be valid?
If the holdout costs the team real renewals, how do you justify asking for it, and to whom?