Duolingo Data Scientist Interview Guide

Duolingo's revenue grew roughly 45% year-over-year to around $748M in 2024, yet the data science org stays lean. Each DS ends up owning a surface area that would be carved across multiple people at a comparably sized tech company, which is exactly why the interview process skews so heavily toward product ownership and causal inference rather than pure modeling chops.

Duolingo Data Scientist Role

You're embedded inside a product squad (monetization, growth, engagement, or new subjects like music and math) rather than a centralized analytics function. Your loop runs from proposing an analysis, through designing the experiment and writing the queries in BigQuery, to producing a written memo with a ship-or-kill recommendation for product leadership. Success after year one means you've driven multiple experiments from hypothesis to launch decision and created at least one metric definition the team now tracks weekly.

A Typical Week

The split that catches most candidates off guard is how much of the week goes to written documents. Duolingo's "Test It First" culture means every experiment gets a formal write-up following an internal template: hypothesis, methodology, caveats, and a concrete recommendation. If you don't enjoy writing clearly for non-technical readers, this role will wear you down fast.

Projects & Impact Areas

Subscription monetization carries the highest stakes: you'll model Super Duolingo conversion funnels, run pricing experiments, and wrestle with revenue attribution when users sit in multiple overlapping A/B tests. New subject expansion (music, math) is a different beast entirely, because you're defining success metrics from zero for learning modalities that behave nothing like language courses. Connecting those two worlds is the causal inference work of disentangling interaction effects across simultaneous experiments, a recurring challenge that touches nearly every product area.

Skills & What's Expected

Expert-level statistics and business acumen are the two dimensions that actually gate your candidacy, but don't underestimate the applied ML expectation. The role requires strong practical ML experience for behavioral modeling and engagement prediction, even though you won't be deploying production model services. What's genuinely underrated in prep: SQL and software engineering quality. DS here are expected to write production-grade queries, contribute to data pipelines (including building metrics and pipelines from scratch in new product areas), and review teammates' code. What you can safely deprioritize: GenAI and modern AI architectures, which aren't part of this role's core requirements.

Levels & Career Growth

Duolingo posts new-grad DS and staff DS as distinct roles with very different scope expectations. The jump from mid-level to senior hinges on owning experiment design end-to-end without your manager reviewing methodology, while the jump to staff means owning the entire analytical narrative for a product area (the staff financial forecasting role in monetization, for instance, shapes investor-facing metrics directly). What blocks promotion most often, from what employees report, is breadth of influence: technical brilliance alone won't get you there if you're not proactively shaping the product roadmap with your analyses.

Work Culture

The hybrid schedule requires time in the Pittsburgh HQ each week, and the culture notes from the team describe most people being offline by 6 PM, so the pace is disciplined rather than grueling. Duolingo's operating principles ("be direct," "ship it") play out concretely in things like the experiment review council, where DS present launch decisions and are expected to flag flawed designs in the room, not days later over Slack. One honest tradeoff: the team is small enough that context-switching between experiment readouts, ad-hoc queries, and pipeline fixes is constant, and long uninterrupted blocks for deep analysis are rare.

Duolingo Data Scientist Compensation

Duolingo's RSUs vest over four years with a one-year cliff, so your equity component carries real risk if you're unsure about staying. Because DUOL is a post-IPO stock with a relatively small float, the market value of your grant at vesting can differ meaningfully from the number on your offer letter. Factor that uncertainty into how you weight equity versus guaranteed cash when evaluating the package.

Base salary and the initial RSU grant size are where candidates report the most flexibility. If you're comparing against offers from larger tech companies, remember that Duolingo's Pittsburgh headquarters means your dollar goes further on housing, taxes, and daily expenses. Don't just match raw numbers; instead, quantify that purchasing power gap and use it to push for a stronger equity grant, which is the component most likely to compound if Duolingo's subscriber growth continues its trajectory.

Duolingo Data Scientist Interview Process

The whole loop takes about five weeks from recruiter call to offer. Six rounds is a lot for a company with fewer than 1,000 employees, and the back half (Product Sense, Case Study, Behavioral) all land during the onsite, so that final week is dense. Block real prep time for each round separately, because they test genuinely different muscles.

The most common rejection pattern is weak coding combined with poor product intuition. The Coding & Algorithms round covers SQL but can also include basic data structures or algorithm problems in Python or R, so candidates who only drill queries sometimes get caught flat-footed. The other trap is the Case Study: from what candidates report, interviewers pay close attention to how clearly you structure and communicate your analysis, not just whether your methodology is correct.

Duolingo Data Scientist Interview Questions

/*
Daily subscription funnel by app-local date.
Assumptions:
- events table has one row per event.
- app_local_date is already computed in the app timezone.
- relevant event types: 'paywall_view', 'trial_start', 'purchase'.
- purchase events for subscription have product_type = 'subscription'.

BigQuery Standard SQL
*/
WITH paywall_user_day AS (
  SELECT
    app_local_date,
    user_id
  FROM `analytics.events`
  WHERE event_name = 'paywall_view'
  GROUP BY app_local_date, user_id
),
trial_starts AS (
  SELECT
    app_local_date AS trial_start_date,
    user_id,
    MIN(event_ts) AS trial_start_ts
  FROM `analytics.events`
  WHERE event_name = 'trial_start'
  GROUP BY trial_start_date, user_id
),
purchases AS (
  SELECT
    user_id,
    event_ts AS purchase_ts
  FROM `analytics.events`
  WHERE event_name = 'purchase'
    AND product_type = 'subscription'
),
trial_conversions_7d AS (
  SELECT
    t.trial_start_date,
    t.user_id
  FROM trial_starts t
  JOIN purchases p
    ON p.user_id = t.user_id
   AND p.purchase_ts >= t.trial_start_ts
   AND p.purchase_ts < TIMESTAMP_ADD(t.trial_start_ts, INTERVAL 7 DAY)
  QUALIFY ROW_NUMBER() OVER (
    PARTITION BY t.user_id, t.trial_start_date
    ORDER BY p.purchase_ts
  ) = 1
)
SELECT
  d.app_local_date,
  COUNT(DISTINCT d.user_id) AS paywall_view_users,
  COUNT(DISTINCT t.user_id) AS trial_start_users,
  COUNT(DISTINCT c.user_id) AS paid_within_7d_users
FROM paywall_user_day d
LEFT JOIN trial_starts t
  ON t.user_id = d.user_id
 AND t.trial_start_date = d.app_local_date
LEFT JOIN trial_conversions_7d c
  ON c.user_id = t.user_id
 AND c.trial_start_date = t.trial_start_date
GROUP BY d.app_local_date
ORDER BY d.app_local_date;

/*
D7 retention by experiment variant.
Definitions:
- Anchor date is the user's first paywall_view while in the experiment.
- D7 retained = at least one 'learning_session' on app_local_date = anchor_date + 7.
- Unit of analysis is user.

Assumptions:
- exposures table: one row per exposure with (user_id, experiment_id, variant, exposure_ts, exposure_app_local_date).
- events table: one row per event with (user_id, event_name, app_local_date).

BigQuery Standard SQL
*/
WITH first_exposure AS (
  SELECT
    user_id,
    variant,
    MIN(exposure_ts) AS first_exposure_ts,
    ANY_VALUE(exposure_app_local_date) KEEP (DENSE_RANK FIRST ORDER BY exposure_ts) AS anchor_date
  FROM `experimentation.exposures`
  WHERE experiment_id = 'annual_paywall_v1'
  GROUP BY user_id, variant
),
d7_activity AS (
  SELECT
    e.user_id,
    1 AS retained_d7
  FROM `analytics.events` e
  JOIN first_exposure x
    ON x.user_id = e.user_id
   AND e.app_local_date = DATE_ADD(x.anchor_date, INTERVAL 7 DAY)
  WHERE e.event_name = 'learning_session'
  GROUP BY e.user_id
),
user_level AS (
  SELECT
    x.user_id,
    x.variant,
    IFNULL(a.retained_d7, 0) AS retained_d7
  FROM first_exposure x
  LEFT JOIN d7_activity a
    ON a.user_id = x.user_id
)
SELECT
  variant,
  COUNT(*) AS users_exposed,
  AVG(retained_d7) AS d7_retention_rate
FROM user_level
GROUP BY variant
ORDER BY variant;

The distribution skews toward questions where you're reasoning about Super Duolingo conversion funnels, streak retention tradeoffs, or paywall pricing, then immediately asked to prove the causal revenue lift of whatever you proposed. That compounding of product sense into causal inference is where most candidates break, because a single interview question can start with "define the right metric for a 7-day trial" and end with "the rollout was by app version, not random assignment, so justify your estimator." The prep mistake that'll cost you: drilling SQL sessionization and ML algorithms while underinvesting in the ability to connect a freemium paywall design decision to an incremental revenue estimate using diff-in-diff or CUPED on Duolingo's overlapping experiment stack.

Practice questions modeled on Duolingo's subscription monetization and experiment design focus at datainterview.com/questions.

How to Prepare for Duolingo Data Scientist Interviews

Duolingo's company strategy frames the ambition plainly: evolve from a language learning app into a broader education platform while keeping the freemium model that fuels growth. Revenue reached $964M (41% year-over-year growth), yet the company still operates with roughly 830 employees, a ratio that tells you something about how much ownership each person carries.

Their operating principles emphasize directness and speed, two values that shape what interviewers actually want to hear. Read both posts before you prep anything else, then skim the investor relations page for the latest subscriber and DAU numbers.

Most candidates blow the "why Duolingo" question by defaulting to "I believe in accessible education." That's table stakes. A stronger answer shows you understand the specific tension the business lives inside: growing subscription revenue without degrading the free tier that drives top-of-funnel user acquisition. If you can connect your DS skills to that tradeoff (experiment design for conversion flows, causal measurement of monetization changes on retention), you're speaking the company's language instead of reciting its mission statement.

Spend at least two weeks using the app daily before your interview. Notice when you hit a paywall, how streak recovery works, and where upsell prompts appear. That kind of product fluency turns a rehearsed answer into a credible one.

Try a Real Interview Question

| experiment_assignments | |
|------------------------|----------------------------------------------------------------|
| user_id | assignment_ts        | experiment_id | variant   |
|--------|-----------------------|---------------|-----------|
| 101    | 2025-01-01 10:00:00   | paywall_v1    | control   |
| 102    | 2025-01-01 11:00:00   | paywall_v1    | treatment |
| 103    | 2025-01-02 09:00:00   | paywall_v1    | control   |
| 104    | 2025-01-02 12:00:00   | paywall_v1    | treatment |

| subscriptions | |
|--------------|---------------------------------------------------------------|
| user_id | event_ts             | revenue_usd | product |
|--------|-----------------------|------------:|---------|
| 101    | 2025-01-03 08:00:00   | 9.99        | monthly |
| 102    | 2025-01-10 10:00:00   | 9.99        | monthly |
| 104    | 2025-01-05 12:30:00   | 59.99       | annual  |
| 104    | 2025-01-06 12:30:00   | 9.99        | monthly |

-- Write a SQL query that produces the requested metrics for experiment_id = 'paywall_v1'.

WITH base AS (
  SELECT
    a.user_id,
    a.variant,
    a.assignment_ts
  FROM experiment_assignments a
  WHERE a.experiment_id = 'paywall_v1'
),
rev_7d AS (
  SELECT
    b.user_id,
    b.variant,
    SUM(CASE
      WHEN s.event_ts >= b.assignment_ts
       AND s.event_ts < b.assignment_ts + INTERVAL '7 day'
      THEN CAST(s.revenue_usd AS DECIMAL(18,2))
      ELSE 0
    END) AS revenue_7d
  FROM base b
  LEFT JOIN subscriptions s
    ON s.user_id = b.user_id
  GROUP BY 1,2
),
metrics AS (
  SELECT
    variant,
    COUNT(*) AS n_users,
    SUM(CASE WHEN revenue_7d > 0 THEN 1 ELSE 0 END) AS n_converters,
    AVG(CASE WHEN revenue_7d > 0 THEN 1.0 ELSE 0.0 END) AS conversion_rate,
    AVG(revenue_7d) AS arppu
  FROM rev_7d
  GROUP BY 1
),
vals AS (
  SELECT
    MAX(CASE WHEN variant = 'control' THEN arppu END) AS arppu_control,
    MAX(CASE WHEN variant = 'treatment' THEN arppu END) AS arppu_treatment,
    MAX(CASE WHEN variant = 'treatment' THEN n_users END) AS n_users_treatment
  FROM metrics
)
SELECT
  variant,
  n_users,
  n_converters,
  conversion_rate,
  arppu,
  CAST(NULL AS DECIMAL(18,2)) AS incremental_revenue
FROM metrics

UNION ALL

SELECT
  'incremental' AS variant,
  CAST(NULL AS BIGINT) AS n_users,
  CAST(NULL AS BIGINT) AS n_converters,
  CAST(NULL AS DOUBLE PRECISION) AS conversion_rate,
  CAST(NULL AS DECIMAL(18,2)) AS arppu,
  (arppu_treatment - arppu_control) * n_users_treatment AS incremental_revenue
FROM vals;

Duolingo's engineering blog makes clear the org values clean, efficient code, and from what candidates report, the DS coding round reflects that bar. Practice both algorithm problems and complex funnel/sessionization queries at datainterview.com/coding.

Test Your Readiness

The questions you'll face skew toward Duolingo's actual business decisions, so product context matters more than memorized formulas. Sharpen that context with Duolingo-specific practice at datainterview.com/questions.