The career changer

Think of a teacher who has spent eight years in a classroom. Every day, she watches people try to learn something hard. She sees who stops coming in October. She sees who is afraid to speak, and who does the homework but never improves. She has never written a confidence interval in her old job, but she has collected a lot of evidence about how people learn.

Many teachers, nurses and other professionals move into data science after a bootcamp. They apply to companies whose product they understand from the inside. For a language-learning app like Parlo, a language teacher is a natural candidate. The usual gap is depth: the bootcamp taught SQL and statistics in a few weeks, and other candidates have used them for years.

Inés Navarro is from Spain. She has lived in Canada for four years. She speaks Spanish, Catalan and English, and English is her third language. On a video call she is clear, and sometimes she stops to find a word. You will see that this does not matter to the interviewer at all.

IN

The candidate

Inés Navarro

Language teacher for 8 years, now a junior data scientist

• 8 years teaching Spanish and English
• Data science bootcamp, 2026
• SQL, Python, Tableau
• Toronto

• Data Science Certificate, Lakeshore Data Academy, Toronto (14 weeks, full time), 2026
• Capstone: analysed open data on 1,900 adult English learners at Northway Newcomer Centre; showed that conversation circles increase the pass rate of the level test by 18%
• Built a Tableau dashboard for the centre’s coordinators, used in their monthly planning
• Spanish teacher, Maple Language School, Toronto, 2022–2025 (adults, evening classes)
• Secondary-school teacher of English and Spanish, Girona, Spain, 2016–2022
• Python (pandas, scikit-learn), SQL (PostgreSQL), Tableau, A/B testing basics

What Thandiwe sees before the call

Thandiwe Mokoena reads the CV next to the job description. The domain fits better than any other CV in the process. Inés taught languages for eight years, and her project is about how adults learn a language. Parlo’s main open question about “Talk with Pip” is exactly this: do people learn more when they talk?

Two things make her careful. First, the SQL screen notes say “passed, but slow on joins; used the AI assistant well and explained it”. Second, the CV says that conversation circles “increase” the pass rate. That word claims a cause. Thandiwe does not assume that it is wrong. She wants to know how Inés knows.

Her short list for the call:

• How did Inés build her numbers, and does she know their limits?
• Does she understand what makes a comparison fair?
• How deep is her statistics? The role needs experiments every week.
• What does her teaching experience add to the Parlo case?

Terms on this page

• Day-30 retention: of the people who joined on a given day, the share who still use the app about 30 days later.
• Weekly active users: people who used the app at least once in a week. What “used” means is a choice, and it matters.
• Percentage rollout: a feature is turned on for a share of users, for example 20%, chosen at random by the experimentation tool.
• LEFT JOIN: a SQL join that keeps every row of the first (left) table. Where the second table has no match, its columns are empty (NULL). See Join tables.
• Confidence interval: a range of values for the true effect that fit the data well. A narrow range means we know more. See Estimate uncertainty.
• Selection bias: the people in a group chose to be there, so they differ from the others in ways that also change the result.
• Power: the chance that an experiment finds an effect of a given size if that effect is real. More users give more power.
• Placement test: a short test that tells a learner which level to start at.

The interview

Minutes 0 to 5: the opening

Thandiwe(interviewer):

Hi Inés! Can you hear me okay?

Inés(candidate):

Yes, yes. Hi. I hear you well.

Thandiwe(interviewer):

Good. I’m Thandiwe, senior data scientist here, I work mostly on onboarding. So, the plan. First you tell me about one analysis you did, and I’ll ask a lot of questions about it. That’s normal, don’t worry. Then I give you a real problem we have right now. At the end, your questions. And this round is just us talking. No AI tools, no searching. Paper is fine.

Inés(candidate):

Okay. Perfect. I have my paper here.

Minutes 5 to 20: an analysis that changed a decision

Thandiwe(interviewer):

So. Tell me about an analysis that changed what someone decided.

Inés(candidate):

Okay. It’s my bootcamp project, but with real data. Northway is a centre for newcomers here in Toronto. They give free English classes. And they publish their class data, without names, for researchers. About 1,900 learners, two years.

The question was from the coordinator. They have conversation circles, it’s like a club, one hour a week, people just talk. Very few people came. And she wanted to know: is it worth it, or do we close it and put the volunteers in normal classes?

What I did is, I compared learners who came to the circles, at least four times, with learners who did not. And I looked at the level test at the end of the term. The ones who came passed much more. So they did not close it. They opened two more circles, in the evening.

What the interviewer is checking

A clear start: a real person with a real decision, the data, and what changed. Thandiwe now goes into the numbers, one level at a time.

Thandiwe(interviewer):

Nice. Your CV says 18%. Is that 18 percent, or 18 points?

Inés(candidate):

Ah. Points. It was… 61% against 43%, more or less. I should write “points”, you are right.

Thandiwe(interviewer):

Okay. And how did you put the learners and the test results together? Two tables?

Inés(candidate):

Yes. One table of learners, one table of test results. I did a LEFT JOIN, learners on the left.

Thandiwe(interviewer):

Let me draw something. She types a small table on the shared whiteboard. Three learners: Ana, Bilal, Chen. Two test results: Ana and Chen. Bilal never took the test. After your LEFT JOIN, how many rows?

Inés(candidate):

Two. Ana and Chen.

Pause.

No, wait. Sorry. That is the INNER JOIN. LEFT keeps all the learners. So three. Bilal is there, and his score is… empty. NULL.

Thandiwe(interviewer):

Yes. And if Ana took the test twice?

Inés(candidate):

Then Ana is two times. Four rows.

Thandiwe(interviewer):

Right. So now the important one. When you calculated the pass rate, was Bilal in it?

Inés(candidate):

Hmm. Let me think. I used the average of the “passed” column. And the average… skips the NULL. So no. Bilal was not in it. The 61 and the 43, they are only people who did the test.

How the interviewer goes deeper

Thandiwe uses a tiny example on the whiteboard instead of a quiz question. Inés makes a mistake, and she corrects herself in a few seconds. Then the real point comes: SQL’s AVG ignores NULL values, so the pass rate counts only the people who took the test. This is not a syntax detail. It changes what her number means.

Thandiwe(interviewer):

And who doesn’t take the test, usually?

Inés(candidate):

The ones who stop coming. In my school also, it was like this. In October you have thirty students, in December you have twenty. And the twenty are the motivated ones.

Oh. So maybe the circle people also… they stay more, because they are more motivated. And I only see the ones who stay. I did not think about this in the project. I don’t know how the 18 changes, honestly.

Thandiwe(interviewer):

That’s okay. Did you do anything about motivation at the time?

Inés(candidate):

A little. I compared inside the same starting level. A2 with A2, B1 with B1. The difference was still there, but smaller. Around ten points, I think, maybe twelve. But I know it is not… how do you say, it does not prove the circles are the cause.

Where this goes wrong

“Increase” on the CV claims a cause, but the data shows a link. Inés does not defend the word. She explains the limit herself. That saves her here. A stronger CV line would say: “Learners who joined conversation circles passed more often (about 10 points more at the same starting level); the analysis cannot show that the circles caused it.”

Thandiwe(interviewer):

Last question on this. If you did it again, what would you change?

Inés(candidate):

I would count the people who left. Like, “passed”, “failed”, and “did not finish”. Three groups, not two. And maybe I ask the centre to try the circles in a fair way. For example, one term, they invite half of the new learners, chosen random, and they see.

Thandiwe(interviewer):

“Chosen at random.” Good. That’s the experiment. Okay, I’m moving on now.

Your English does not need to be perfect

Inés searches for words, says “chosen random”, and once asks “how do you say”. Thandiwe never comments on it. She listens for the thinking, and the thinking is good: Inés found the selection problem herself, once the example made it visible. If English is not your first language, speak simply and take the time you need. See Your English does not need to be perfect.

Minutes 20 to 26: the Parlo case, first part

Thandiwe(interviewer):

Okay, now our real situation. Five weeks ago we launched “Talk with Pip”, AI conversation practice, to 20% of new users. Weekly active users in that group went up about 6%. But day-30 retention for new users went down from 31 to 29%. Our product manager wants to give it to everyone before January. If you were our data scientist, what would you do?

Inés(candidate):

Okay. First some questions, if I can. The 20%, how did you choose them? Random?

Thandiwe(interviewer):

Random, by the experimentation tool, a percentage rollout. But only on iOS. Android wasn’t ready.

Inés(candidate):

Only iOS. Okay. She writes. And the 31%, who is it?

Inés learns that the 31% is Kenji’s dashboard: all other new users, on both platforms. She asks if iOS and Android users are the same (“Maybe iOS people pay more?”). Thandiwe says that iOS users stay longer, so Inés asks for iOS users without Pip. The answer is about 33%, so the fair gap is about 4 points, not 2. “Hmm. Now I am more worried, not less,” she says.

Minutes 26 to 38: how many people, and which people

Thandiwe(interviewer):

Laughs. Okay, remember that. How many people do you think are in that 29%?

Inés(candidate):

Sorry, can you repeat? The sound cut a little.

Thandiwe(interviewer):

Sure. The 29%. How many people is that, do you think?

Inés(candidate):

Ah, okay. Five weeks… I don’t know how many new users you have. Many thousands, I guess?

Thandiwe(interviewer):

About 2,400.

Inés(candidate):

Only 2,400? But it is five weeks, and 20%…

Oh, wait. Day 30. You need thirty days to know. So only the people from the first days have the number. The first week, more or less.

Thandiwe(interviewer):

Exactly. The first five days of sign-ups. End of August.

Inés(candidate):

End of August. You know, in a school, September is when everybody starts something. New courses, new gym. And in October, many people stop. January is like this too, with the new year. But they are not the same people. So I am careful with only one week.

How the interviewer goes deeper

Here Thandiwe helps a little, and Inés finds the point herself. A day-30 number after five weeks covers only the users who joined in the first five days. They are a small group, and they all joined in the back-to-school week. Inés adds something from her classroom: people who start in September may behave differently from people who start in January.

Thandiwe(interviewer):

So, iOS against iOS, 33 to 29, with 2,400 people in the Pip group. Is the drop real?

Inés(candidate):

I can’t calculate it in my head. I know there is a test for two proportions, we did it in the bootcamp. I would look up the formula, or use Python.

Thandiwe(interviewer):

Fair. I’ll give it to you. The 95% interval for the difference is about minus 6 to minus 2 points. What does that tell you?

Inés(candidate):

Minus 6 to minus 2. It does not touch zero. So it is a real drop, I think. Maybe small, 2 points. Maybe big, 6.

Thandiwe(interviewer):

So would you tell Kenji, “Pip lowers retention by 4 points”?

Inés(candidate):

For iOS people who joined in that week… yes, probably. They were chosen random, same week, same phone. But for everybody in January? Android, other countries, new year people? I don’t know. One week in August is not January.

Why this works

Inés cannot compute the interval, and she says so plainly. But she reads it correctly, and she separates two questions. Is the drop real for the people in the test? Probably, because the tool chose them at random. Will it be the same for everyone in January? The data cannot say. Many people with more statistics training mix these two up.

Minutes 38 to 50: the Parlo case, second part

Thandiwe(interviewer):

Okay. What else would you want to look at?

Inés(candidate):

The question I really want to ask is: do they learn more? Because weekly active is… people open the app. Retention is people stay. But Parlo sells learning, no? Do you measure if they speak better?

Thandiwe(interviewer):

Pause. Honestly, not really. We have a short placement test. But few people take it after the first day.

Inés(candidate):

Then it’s the same problem as Northway. Only the motivated people take the test.

What I would do: take a random sample from both groups, Pip and no Pip. Maybe 500 each. And invite them to a short speaking test at day 30. Give something for it, a free week of Plus. And then you compare. Random people, not the people who want to do the test.

Why this works

This is the best question in the case, and it comes from her teaching years. Parlo measures activity and retention, but its promise to learners is learning. Inés also applies the lesson from her own project. If only volunteers take the test, the result is biased, so she invites a random sample. One open question remains: not everyone invited will take the test. But the start is right.

Thandiwe(interviewer):

I like that. We have never done it. And one thing you didn’t ask: lessons completed went down in the Pip group.

Inés(candidate):

For a teacher, that is not bad by itself. Talking is maybe more useful than the drills. But talking is also hard. For beginners, it can be scary, you feel stupid. My adult students, the first weeks, some of them did not say one word. So I would look by level. Maybe Pip is great for B1 people and too hard for beginners, and the beginners are the ones who leave.

Thandiwe(interviewer):

We do have the level from onboarding. Nobody has split it that way yet.

Inés(candidate):

And also… sorry, one more. Where are the people? On the bus, in the metro, you cannot speak to your phone in a loud voice. In some countries this is very normal, in others no. So maybe some people never try Pip.

Thandiwe(interviewer):

That’s real. In some countries a lot of people skip it. Okay. Now let’s say Kenji agrees to do a proper experiment. Random new users on both platforms, half with Pip. How many users do you need, and how long?

Inés(candidate):

Hmm. This is the power, right? I know the idea. I would use a calculator. But I need… what difference is important for you. One point? Two points?

Thandiwe(interviewer):

Say we care about one point of day-30 retention.

Inés(candidate):

One point is small. So many people. More than 2,400, for sure. Much more. But how many… I really don’t know. Sorry.

Thandiwe(interviewer):

It’s fine. And how long?

Inés(candidate):

At least thirty days, because of day 30. Plus the time to get the people.

The edge

This is the edge. Inés knows the ideas: power, the size of the difference, the 30-day wait. She cannot turn them into a number, and the role needs this every week. She does not invent a number, and that is right. But she also stops too early. She could reason her way to a rough size.

A stronger answer

A stronger answer does not need the formula from memory. It needs a rough estimate and the reasoning behind it:

“I don’t remember the formula well, so let me try roughly. Retention is about 30%. The difference we care about is 1 point, so that’s very small. I remember that small differences need tens of thousands of people per group. I think it’s around 30,000 per group for one point, but I would check it with a calculator before I say it to anyone. How many new users do we get a day? If it’s a few thousand on both platforms, we need about two weeks to get the people. Then we wait 30 days for the last ones. So about six weeks. If we start this month, we know before January.”

Parlo gets about 5,500 new users a day, so the six weeks are right. With iOS only, it would take about two months. The estimate also helps the decision: Kenji can have a fair answer before the January peak.

Thandiwe(interviewer):

Okay, last part. Imagine Kenji sends you a message: “Do we give Pip to everyone, yes or no?” What do you tell him?

Inés(candidate):

I tell him: not yet. People use the app more with Pip, that is good. But when we compare iOS with iOS, retention is about four points lower. It’s only one week of people, so I am not sure for January. But it is a warning. First we find out why. Maybe the beginners, maybe the lessons. And we test if they learn. Because if they learn more, maybe a small drop is okay. If they don’t learn more and they leave more, then no.

Thandiwe(interviewer):

And if he says, “I can’t wait, January is our biggest month”?

Inés(candidate):

Then… maybe give it first to people who are not beginners. And keep a small group without Pip, to keep measuring.

Why this works

Her answer is short, honest about what is not known, and in Kenji’s language. She gives the bad news clearly, and she does not make it bigger than the data allows. When he pushes back, she offers a compromise: start with learners who are not beginners, and keep a group that does not get Pip, so the team can keep measuring. Thandiwe notes that this is a “long-run holdout”, even though Inés does not use the term.

Minutes 50 to 60: Inés's questions and the close

Inés(candidate):

I have two questions. First one. Who at Parlo works on learning? Like, the science of learning?

Thandiwe(interviewer):

Good question. There are two linguists in the content team. They design the lessons. To be honest, data science doesn’t work with them much. We talk about retention, they talk about learning. It’s a gap.

Inés(candidate):

Okay. And the second. If I start, what is the first three months? What is a normal week?

Thandiwe(interviewer):

You’d own one product area, alone with a product manager. Many small questions from people, so SQL every day. And almost every week, an experiment to design or to read. The end of each quarter is busy.

Inés(candidate):

Okay. Thank you for saying it directly. The statistics part, I know I need to grow.

Why this works

Both questions get true answers that are not in the posting. The first shows a gap at Parlo that matches her strength. The second shows the real weekly work, so Inés can judge the fit too. See An interview is a conversation in both directions.

Thandiwe(interviewer):

Thanks, Inés. I really liked the learning question. Mira will contact you this week about next steps.

Inés(candidate):

Thank you. It was… I enjoyed it, really. Bye!

After the call: Thandiwe’s notes

Candidate: Inés Navarro. Product case, 60 minutes.

Strong evidence

• Product sense for a learning product. Asked whether learners learn more. Proposed a random sample with a speaking test. Nobody else has asked this.
• Fair comparisons. Asked for iOS against iOS and found the 4-point gap. Separated “real for these users” from “true in January”. Found the selection problem in her own project once we looked at the NULLs.
• Communication with a PM. Clear, honest recommendation; a sensible compromise when pushed.

Some evidence

• SQL. Understands joins after a small example; first answer on LEFT JOIN was wrong, then corrected herself. Same picture as the SQL screen.
• Reading a confidence interval. Read it correctly, could not compute it.

No evidence

• Experiment design in numbers (power, sample size, duration), causal inference methods, machine learning.

Concern

• Statistics depth. Reached her edge at sample size. This role designs experiments alone with a PM every week. She would need months of close support.

Recommendation: no for this role, with regret. But Simone, please talk to her about the learning-outcomes problem. She saw in ten minutes what we have not measured in a year.

What happened next

Simone read the notes and called Inés the next week. She was honest. The data scientist role needs more statistics than Inés has today, and the one vacancy went to another candidate, Chloe Martin.

But Simone had a problem that Thandiwe’s notes described well. Parlo could not say whether learners learn. So Simone created a six-month analytics contract on the learning-science side. Inés would work with the two linguists and Thandiwe. Her first task is to design the speaking test for a random sample, and to build the data for it. The contract is not a permanent role, and Simone said so clearly. She also said that a permanent role may follow if the work goes well, but she did not promise it.

Inés accepted. In her first month she asked Thandiwe for one hour a week on statistics. Thandiwe said yes.

This is a common way into a data team for career changers. You may not get the role you applied for. But if you show a real strength in the interview, a team can sometimes create a smaller role for you. That is partly luck, and partly what you showed.

What to take from this

• Your past work is evidence. Inés’s best answers came from her classroom: September starters, scared beginners, learners who stop coming. See Your experience is a head start.
• Correct yourself out loud. She said “two rows”, then “no, wait, three”. A quick correction is a good sign, not a bad one.
• Know what your numbers mean, not only how you got them. The NULLs in her LEFT JOIN changed who was in her pass rate. That was the real lesson of the warm-up.
• Say what you don’t know, then reason as far as you can. She was right not to invent a sample size. She could have gone further with a rough estimate.
• Your English does not need to be perfect. The interviewer judged her thinking, and her thinking was clear.

If your profile is like Inés’s

Your domain knowledge is rare, and it shows in a case. The gap is statistics and SQL depth, and you can close a lot of it in a few weeks of focused practice.

1. Practise joins with tiny tables on paper. Three rows on the left, two on the right, one duplicate. Predict the result of each join, then run it. Then compute an average and check which rows it used. See Join tables and Aggregate correctly.
2. Learn to estimate a sample size roughly. Use a power calculator for a few cases: 30% to 29%, 30% to 28%, 10% to 9%. Write down how the numbers change. After an hour you will have a feeling for the size, and you can say it in an interview. See Compare alternatives and Estimate uncertainty.
3. Fix your project’s weak point, and write it in the README. For Inés: add “did not finish” as a third group, and say what the analysis can and cannot show. An honest limit makes a portfolio stronger. See Portfolio anatomy.
4. Rewrite your CV verbs. “Increased” claims a cause. “Was linked to” describes what you found. Check every claim like this.
5. Practise the Friday answer. Take any case and explain your recommendation in four sentences to a friend who does not work with data. See Explain decisions.
6. Apply where your domain matters. Education, health, language and training companies need people who understand their users. In those interviews, your past is an advantage.