Shadowing Practice: FULL Data Science Mock Interview | SQL, Statistics, Business - Learn English Speaking with Video

In this video, you're going to watch a data scientist work through a technical on-site round that includes SQL, stats, and a business case.

Johnny, right?

Yes, Johnny.

Nice to meet you.

How's it going?

Nice to meet you.

You're not related to Johnny Bravo at all, right?

Not at all, though.

Yeah, I can't say I haven't watched a show before.

I did watch it, definitely.

You got the same haircut and everything.

It was like the first thing I noticed.

Yeah, there's definitely some similarities.

Cool.

Well, thanks for taking the time to do this interview.

Just wanted to start off with a little bit of logistics, just to give you a groundwork for kind of what we're going to go through during this time.

We're going to be doing the interview for 45 minutes.

It's going to be a technical round and there's going to be three sections, a SQL section, a stat section and a business use case.

that we're going to walk through together.

Most of it will be more conversational in nature, though in the beginning, the first section is going to be the SQL section,

and we will go through a little bit of coding.

So what I'm going to do is for that section, just give you a prompt and if you can share your screen

and just pull up a simple notepad and show me kind of how you would write things.

It doesn't have to be a specific SQL syntax or anything, but just pseudo-SQL, if that's a thing, like that would be good enough.

So any questions before we get started?

No, I'm guessing from the pseudo-SQL part, we're not going to like,

obviously no AI, no chat GPT on the side and also no reference docs.

Is that correct?

Yeah, that's a great question.

I think

if you need like a specific function that's pretty like you

know um you know like we want to simulate the work

environment as much as possible uh barring the chat gpt ai stuff

so um i think within reason you can definitely

but um feel free to let me know if you're going to look something up that would be helpful sounds good

okay sounds good all right

so we'll let's start with the first one you're working with the right share company

and you have access to the following tables and here's a snippet just to to be able to look at each table.

So that is the first table.

OK.

Let me know if that looks too weird.

That should be OK.

I'll go ahead and share the screen with the notepad.

Sounds great.

And here is the second table.

OK, cool.

So we've got two.

All right, let's see if we can format it a little bit.

It's like we're going to be fighting this a little bit.

But in general, so I see we've got this.

Okay, so trip, user ID, driver ID, city, distance.

And the prompt over here, we are a rideshare company.

That's right.

Cool.

Okay.

This...

Cool.

Okay.

And then the second table, shorter, thank goodness.

just three columns

cool okay user id sign update um user type and

so just understanding here so i know trip is your the trip driver is the the one who's working for Uber.

And so user here is really our customer.

That's right.

Okay, cool.

Yeah.

So yeah, you kind of nailed it on the dot.

There's two tables here.

One that shows the trip data and some of the metadata around it.

And then one is the users.

And actually, I want to clarify a little bit.

Let me give you another row of that user table.

Okay.

because the user type is relevant here.

It could be the writer, but it could also be the driver in that user data as well.

So you can feel free to paste in a second row.

Yeah, and thanks for doing the formatting.

I think it's actually helpful for me as well.

So I appreciate it.

Cool.

Great.

So here's the first question.

I'll paste it in as well so that you can have it while we're going through it.

So the first one is write a SQL query to calculate the average trip distance per city.

So obviously you don't have the whole data set here that you're looking at, but if you can just write a query that gets at,

you know, how you would calculate that.

Yeah, cool.

Okay, so when I write SQL, I immediately go, I know there's a select,

there's my from, and then we'll have some other group buys or wears.

So we're talking about calculate average trip distance.

So that's immediately like telling me, all right, so I want the average trip distance.

It looks like it's provided here on the trip data.

I'm going to make an assumption that the table name here for that is trips and the table name here is users.

Is that okay?

that yes that was a copy and paste there yes very good cool

so um all right so i want average trip distance

so we've got distance kilometers uh which is going to be from our trips table and

per city so per city if we're doing an average we're gonna need a group by we can do this

group by city here.

100

So I believe this should give us the average distance of trips by city.

101

And I guess we can add the city up top so that we know which distance connects to what city.

102

Okay, great.

103

Yeah, that looks good to me.

104

All right, let's move on to the next one.

105

Find the top five users who are writers.

106

We're talking about writers in this situation who spent the most in July of 2023, and be sure to include their total spend as well.

107

Okay.

108

All right, so we're thinking top five users who spent the most.

109

Okay, so explicitly riders here.

110

In July of 2023, it's like our time frame, include total spend.

111

So our select statement, so we eventually want rider,

112

riders, which is probably going to be users.

113

Yes, I just want to clarify that because it wasn't clear in the schema there.

114

User ID, it could also be rider ID.

115

That's what should connect to that user ID column in the user data table.

116

Gotcha.

117

Okay.

118

So why don't maybe let's start a little simpler.

119

Let's go with just select writers who spent money in July 23.

120

So if I want writers, so let me just select,

121

let me find all users who had a trip in July 2023.

122

Now this July 2023 data is in our trips,

123

so I'm going to need to join here into trips on maybe like trips.usrID.

124

and I should probably just make this a little bit more easier.

125

Sorry, that was a pretty quick jump down.

126

All right.

127

So from users, left join trips on trips user ID equals, so we are the user table, so users user ID.

128

So now we've got users trips.

129

So we can now say where – so I wanted to do July 2023.

130

So we've got trip dates here.

131

So we should be able to do trips.tripdate maybe less than – we are at 2023.

132

2023, so I want to be within 2023.071.

133

Pinterest.tripdate.lesson.081.

134

If I knew how many days were in July, I would use 731, but I don't remember if it's 30 or 31 days.

135

No, that seems perfect.

136

Okay, cool.

137

And one more thing here.

138

So users could be riders or drivers.

139

So we should probably make sure that we're always a driver.

140

Oh, sorry, rider.

141

Cool.

142

so so far looks like we've got users trips in july um

143

so we still need total spend so we just want top five um so that's that's a pretty easy limit five there

144

and we want total spend

145

so spend is probably price trips up price USD we want the sum of that so trip stock

146

cool so we're going to sum this up but in order to do an aggregation we'll need a group by. So let's group by.

147

We want user ID at the end of it.

148

I don't know if it'll get confused between the two tables user ID, so I'm just going to go with users.

149

All right, let's read through this one more time

150

so find the top five so that's limit five users users

151

or writers who spent the most we've got prices in july 2023

152

all right i think i think that's my final answer okay

153

yeah it looks great uh one question i had is is what you think those first five rows would look like

154

if you were to execute this at the moment.

155

What they would look like?

156

Well, I would have a user ID and the total spend on the,

157

for that user yeah for that writer oh

158

but we're looking for top five yeah well top five good good call

159

so that would be i guess we need an order by

160

and we want top and we want to do it by this column

161

so let's do ads maybe trip uh let's do total spend and then we can order by total spend sending

162

awesome all right yeah

163

that looks really good awesome thank you um we'll quickly work through um one more question

164

and this one just like let's um maybe we can talk through how we would do this

165

and you can fill in and write as much code as you can

166

but we have about like a minute more for this section

167

so we'll try to burn through this um for each city

168

rank drivers by total earnings in august 2023 uh okay

169

so you said we had a minute so i probably won't focus on coding too much and more on

170

explaining what i'm thinking so you want drivers by total earnings for each city so i'm thinking

171

we still have drivers we're gonna do a left join on trips to get our

172

city data and then we'll want to group by probably the cities and the drivers Yeah,

173

because this is kind of our tuple right here for each city drivers by total earnings.

174

And I noticed this rank here.

175

So I'm not too familiar with this,

176

but I'm thinking maybe like a rank by rank over partition with city.

177

Yeah.

178

And it looks like we're doing it over the total earnings.

179

So maybe I'll wear like some sum of total earnings.

180

Yeah.

181

Okay.

182

Yeah.

183

That sounds good.

184

That sounds like a very sensible approach to me.

185

So yeah, sounds great.

186

Awesome.

187

Okay.

188

Thanks for running through this portion.

189

I think writing code in an interview setting is never fun.

190

So thanks for humoring me through that.

191

We're going to move on to the next section.

192

And this, I will continue to paste questions to you in the, so that you can have access to it,

193

but you won't code as much.

194

So we're going into more of the stats section here.

195

So here is the prompt just so you started started um

196

and i'll i'll walk through this um so

197

that we can have it together

198

so did you uh right before you start should i will i need the notepad should i just like close

199

that yeah you can feel free to close it yeah okay cool yeah

200

okay so um here's the scenario um you're running an a b test for an online ad platform to evaluate

201

a new bidding algorithm so there's already one um in place

202

and you're evaluating this new one version b against the current one version a

203

and comparing on the two rates so this is ads so um just the rate at which you know

204

a user is looking at something and then whether or not they've clicked

205

so after running the test for a few days the results

206

are as follows group a is the control group it shows 100

207

and 1200 clicks out of 60 000 impressions uh a two percent click-through rate

208

and then group b is the test group had 1350 clicks out of the same number of impression 60k

209

at a 2.25 click-through rate so um we see that just by you know uh just by absolute comparison

210

there's a 0.25% increase.

211

So the marketing team is eager to roll out the new algorithm immediately.

212

And so here are the questions that we want to walk through regarding this.

213

This is mostly a discussion meant to kind of talk through some of the statistical concepts.

214

Okay.

215

First one, I wanted to focus on the central limit theorem here.

216

Why would the central limit theorem

217

be essential like how how does the central limit theorem play into and now analyzing these results

218

yeah so um so the central limit theorem uh as i understand it allows us to

219

take results from a sample and uh say and it kind of gives us this distribution

220

or it allows us to take the results of a sample and apply it on a normal distribution, even if the total population might be skewed.

221

So taking that knowledge and for these particular results, it looks like the test was run on 60,000 impressions,

222

which is a pretty good sample size, I think, unless you're like Amazon or something with millions.

223

but hopefully this is actually a big enough sample size

224

that we can we can now apply a normal distribution to this

225

and therefore make statistical decisions or run statistical metrics such as the z-score standard error

226

or etc okay yeah great awesome um and so regarding those statistical um

227

tests um that kind of segues well into the next question how would you determine whether this difference

228

that we're seeing is meaningful enough to justify rollout like how

229

would you just talk me through like what kinds of things you would look into

230

and um how

231

that would play into you know how the results of those things would play into informing the recommendation you might make

232

yeah i think um well we want to uh identify the statistical significance um so 0.25

233

is i guess our absolute lift in this case

234

and what we'd want to do is we're trying to define figure out the p value

235

and given some alpha such as like 0.05 if our p value given these inputs is much smaller than 0.05,

236

that would tell us that, hey, this test is actually statistically significant.

237

So that would be one sort of check mark I'd want to make sure of.

238

the other thing is on the cost side so even

239

if we've established that this is statistically significant we'd also want to make sure

240

that if we do deploy our new model that

241

it kind of the profit is going to be greater than the actual cost of deploying the model

242

and there are some metrics that we can use to compare okay yes

243

that sounds great um

244

so going back to the statistically significant um portion um how

245

would you help a marketer like let's say marketer has very

246

very like thin statistical background how would you help them understand like what

247

that means like statistically significant and the p-value and the 0.05 like

248

could you could you help create could you help provide like an intuitive understanding of why

249

that leads to um like hey this difference actually is meaningful That might be tough.

250

My understanding is that although we're seeing changes,

251

how would I frame it?

252

So what we're seeing is changes in a very small sample,

253

And we want to make sure that these changes will scale to the larger population.

254

And there are some times where we may see like a 0.25% in a small population.

255

But when we actually deploy it into the larger population,

256

that number could actually drop and be closer to 0.1 or 0.01 and we want to make sure by using the statistics,

257

using these central limit theorem,

258

these confidence interval tests, we're doing our due diligence to make sure that this 0.25% that we see in the sample,

259

it will be the same 0.25% or better in production.

260

Okay, yeah.

261

So if I'm understanding correctly, like the statistical significance, it points to some indicator that

262

that difference is real when we deploy it to the greater population

263

is that how is

264

that a good way of understanding kind of what you're saying there yes yeah okay okay yeah

265

that sounds great um

266

and then going into the cost side um how would you like how how would you um put

267

that into metrics um if you were to

268

so i think it was a really good call out

269

that you have to consider the rollout costs relative to the

270

potential lift it would bring um what what numbers would you

271

consider bringing to this marketing stakeholder group um as you're bringing

272

this analysis to them as you're talking about the rollout plan to them yeah um something

273

things that I might start thinking about, hey, what is our ROI?

274

So right now we have 2% CTR, which is click-through rate.

275

How does that translate into profits?

276

Right now it's just a number.

277

How would we translate that into a dollar amount?

278

So we might have some data about,

279

well uh maybe one percent click-through rate gives us some net profit of like a thousand dollars and

280

so an extra dot to five percent is like all right

281

we're getting i don't know the math like two hundred dollars let's just say uh

282

or it's probably 250 that we're getting so using that we might calculate our roi our return on investment

283

Another, that's definitely one metric.

284

The other metric I want to say is MDE,

285

which I'm forgetting the meaning.

286

Minimum detection effect?

287

Is that the one you're referring to?

288

i believe so yeah minimum detection um the definition is escaping me right now

289

apologies no no worries no worries yeah i think the roi

290

is a great call out i think it's a good place to start

291

and it's it's something that the marketing stakeholders will definitely understand

292

so i think it's a really good place to kind of like um um bridge the gap yeah so Okay, that sounds great.

293

Okay, so let's move on to the third question in this section, which is, what would break our ability to do this analysis?

294

So this is pointing back to the central limit theorem concepts that we had just talked about.

295

As you had mentioned, it kind of underpins our ability to do an analysis like this to begin with.

296

So if you could help me to understand, in what circumstances would we not be able to use those assumptions that we laid out?

297

Yeah.

298

I think there's a few items I'm thinking about.

299

One of them is the segmentation.

300

if we were to find out that, hey, we didn't have a good distribution of segments.

301

So for example, maybe we only got PC users on our A side and device users on our B side.

302

That's a clear skew, which would break that central limit theorem.

303

That's the one my mind goes to.

304

Sambo.

305

There's another one early peaking.

306

So this idea that the more times, if this is maybe the third or fourth time we're running this test,

307

this actually increases the likelihood that maybe this actually increases the likelihood that there was a false positive.

308

And the false positive being that out of the four times we ran it

309

and in the example the fourth time showed a significant like a statistical significance

310

and we want to avoid that because the other three were normal

311

so I'd probably start with those two yeah yeah those are really good considering ad as the industry here.

312

Are there any things that you can think of

313

that would create some potential weirdness or things that are maybe normal about the ad industry,

314

but would introduce gotchas and things to watch out for as we look at this data?

315

Yeah.

316

Okay.

317

So ad bidding algorithms. So probably

318

i think of ads and i just immediately think politics like

319

so there are certain very charged subjects uh that would um that could like skew the results one way or the other

320

versus a more kind of stable topic such as i don't know chicken eggs like who doesn't who doesn't hate chicken eggs.

321

So that could be one example where like now we've got this, people already have their own biases about the ads.

322

And so they might choose A or B based on their own personal biases instead of the actual test.

323

Yeah.

324

And the last question I wanted to follow up about this one is, What questions might you want to ask about how the data was sampled?

325

When you think about the control in the test here, there conveniently hasn't been given to you many details about how the data was sampled.

326

So what questions might you want to ask just to make sure you're covering your bases?

327

Yeah.

328

So in terms of sampling, there's a fancy metric, I think, called SRM.

329

But try not to dive into that.

330

Probably I'd want to ask, like, hey, so it's ad bidding on top of this.

331

I'd want to know kind of how they decided.

332

Was it a random sample?

333

Were these segments or were these divided on a particular customer segment?

334

How was the sample actually recorded?

335

Like, did we actually, when we were recording it, was there a network outage in between?

336

Was there some kind of external economic indicator or something that could have possibly

337

skewed the results during the case of the test.

338

Yeah.

339

So what I'm hearing from there is like, it sounds like you want to really make sure that the two samples weren't taken differently because of some situation,

340

some aberration that could have happened.

341

Yeah, I think those are really good ways to think about that. So awesome.

342

Okay, great.

343

We're making good progress on time here.

344

So we're going to move on to the next, the last section, the business case analysis.

345

So we're going to switch gears.

346

Yeah, we've been kind of switching gears each time.

347

But this one, we're going to go to food delivery.

348

So now you're a data scientist at a food delivery company.

349

And over the past quarter, you noticed a 15% drop in order volume in your largest market.

350

so leadership understandably wants someone to investigate and they they've asked you

351

and they've asked you to investigate and to recommend actions

352

so you have access to customer orders delivery times ratings marketing spend data you may

353

or may not use all this data

354

but that's what you have at your disposal at the moment

355

and so this this one is fairly open-ended

356

and again like would love to just have discussions around it and feel free to ask any questions for clarification.

357

And so we'll start with a very open-ended question, where would you start and why?

358

And yeah, feel free to just, I guess, put to words and articulate as much of your thought process as possible.

359

Yeah.

360

So, I mean, I hear 15% drop and I'm thinking about, am I going to lose my job soon?

361

but we've got the uh obviously that is a huge marker something

362

that i'm thinking about did this happen all at once was it a sudden event

363

or did this happen gradually over time i'd want to

364

probably focus on given that time nature looking at temporal data so for example

365

an easy one might be hey what does my delivery times look like over time what is my

366

looking at marketing spend data

367

so maybe did we have changes in marketing spend has it

368

been relatively similar I'd also might want to compare it not just

369

against I might also want to compare these trends against the same time period last year to see if like,

370

hey, maybe this is kind of expected for a business.

371

Yeah.

372

Yeah.

373

Okay.

374

That's great.

375

So let's, let's focus in on the temporal data.

376

So what hypotheses would you want to form as you think about this, this particular area?

377

yeah so uh well we're thinking food delivery um

378

so i i almost uh want to say has there been a change in uh drivers maybe

379

so the order history uh

380

so have the ratings kind been going down are we seeing a correlation between orders

381

and ratings that would be one of them marketing spend could be huge i mean

382

that is exactly how we get customers onto our platform are we losing like conversion rate there

383

and the third one is i kind of want to stick with seasonality

384

or in general some sort of external factor

385

and maybe we could look at like i would say like maybe customer order trends by year and like compare them

386

and see are we seeing the same sort of trends okay yeah um okay

387

so So let's just deep dive a little bit into one of these areas.

388

So the first one you mentioned, the change in drivers, looking at order history and ratings.

389

What are some of the, like, if you were to form like a testable hypothesis, where would you go with that?

390

you know, trying to figure out some, trying to detect some change in drivers using order history and ratings.

391

Yeah.

392

Drivers order history rating.

393

So, and we're trying to track order volume drop.

394

So driver history.

395

Something I think about is maybe there's a, subsection of drivers that is causing orders

396

or users who have ordered from them to to not order again

397

so because of because of an order

398

that came in from a particular driver will the user order again

399

or will they not so

400

that could be one sort of testable hypothesis system okay okay yeah um

401

and then depend uh so how would your strategy

402

or how would your thoughts on these things change

403

if you saw a sudden drop in 15 percent um in

404

this largest market versus is like a gradual drop over a quarter, let's say.

405

Yeah.

406

If it was sudden, I'm thinking there was probably a flip that was switched.

407

I think more like a flip that was switched,

408

like maybe there was an update that happened right on that time because we're thinking about like, hey, it's a food.

409

I'm presuming an app here maybe a website i'd track it to a singular event versus if it's something gradual

410

um i'm thinking more uh long larger trends maybe there was a tick tock someone made

411

and it's slowly blowing up and as these more and more users are seeing it it's affecting their ordering habits

412

or maybe there's an economic factor like uh i don't know maybe the stimulus checks have stopped

413

and so people just aren't ordering as much food um

414

so yeah okay great great um and yeah

415

so when it comes to like the gradual kind of seasonal trend changes that are happening,

416

let's say we take the hypothesis

417

that you mentioned of maybe there's a subsection of drivers that are causing orders to not happen again because of something.

418

Maybe it's a bad experience.

419

Maybe it's...

420

Delivery times.

421

delivery times maybe yeah maybe it's there's something there

422

and you see a correlation between that and uh ratings um so

423

that what is the metric like let's get really granular let's

424

try to get as granular as we can what is the actual metric you would want to calculate

425

when you are trying to test this hypothesis right so we're thinking like some sort of success metric um

426

So the user, so subsection of drivers and our success is whether a user actually orders again.

427

So I think, I don't know what the name of it would be, but we'd specifically be testing,

428

hey, does the percentage of users that order again, given a particular driver?

429

okay and that would be all time so if a user how like how much time can elapse for a particular

430

user to order again uh well we should probably uh define like some baseline

431

so before this period what was the average time between orders and then use that as the interval in our metric.

432

Yeah.

433

Yeah, that makes sense.

434

And when you say given a particular driver, so percent of users that order again given a particular driver,

435

so is this one metric for each driver or how would you aggregate if you would?

436

Yeah, so I'd want to be on the per driver level,

437

but obviously there's a number of drivers in the database.

438

So we would need a way to kind of compile this metric so we can get per average driver.

439

um i want to say average as well

440

if we average the value well

441

so we have average there's also uh median is what i'm thinking of uh

442

and i'm trying to see like hey what would make more

443

sense i guess i'd have to know a little bit more about the data to make a decision about median or average.

444

Yeah.

445

So between median or average, do you have something like a criteria in mind for when?

446

So how would the data need to look for you to use average versus median, would you say?

447

Yeah.

448

So if the distribution of the per driver metric is closer to normal,

449

normal, I'd want to use average versus if it was skewed.

450

And especially if we have large outliers, I'd want to use median instead.

451

Yeah.

452

Yeah.

453

That sounds good.

454

And then just maybe one last thing there is what

455

if the distribution looked like a really long tail where you have a couple of

456

drivers who are taking up like 80% of the orders

457

and then like the average driver is taking like very few orders okay

458

so in the case that in the case that there's a subsection

459

so there's we could rely on median uh we could also decide to

460

chop off some of those outliers just outright in order to get a better performance.

461

We'd have to be careful about what type of error.

462

We don't want to cut off too much, but we might have a threshold

463

so that we're not polluting our data and our statistics with this group that maybe only happens 5% of the time.

464

Yeah.

465

Okay.

466

Yeah.

467

And then last question I want to ask here is, so that seems like a good metric overall to use.

468

Are there any like guardrail metrics or any like counter metrics you'd want to measure

469

while looking at this just so that you're not,

470

you're kind of being cognizant of any potential tradeoffs that the results of this analysis can cause?

471

Yeah, as we go through this data, I'd want to make sure that we're not affecting latency, for example.

472

As we're doing these tests, we want to make sure we're not adversely affecting the number of orders coming in.

473

We don't want to start the test and orders don't dive even further.

474

So those are like two examples of guard worlds where we don't want to affect our main product

475

while we conduct this test.

476

Yeah.

477

Yeah.

478

So it sounds like the experience of the product

479

and then making sure that the business is staying stable while we're looking through.

480

okay yeah sounds good okay yeah um so those are all the questions

481

that i had for you today um so thank you

482

so much for taking the time to do this interview

483

and uh yeah it was really nice to meet you yeah no i appreciate it thank you that's that's the

484

that was the mock interview um uh uh

485

Khalid what did you think about the candidate how was their performance to you uh i was okay It was definitely,

486

I think, they had a strong start in the beginning.

487

Kind of as we started getting into more of the statistics side, I felt there was a bit more of a general,

488

like a high-level view.

489

And maybe there was a little bit of struggle.

490

I know he wasn't able to go into MDE a little bit.

491

But something to note there, I don't know how that'll play in the larger role.

492

Yeah.

493

Johnny, what do you think?

494

Does that track what Khalid's saying?

495

Yeah, I would totally agree that the candidate was really strong in the beginning with the sequel.

496

I think he did a really great job of different aspects of that section overall.

497

I think so that I can kind of dig into it section by section since.

498

Yeah, that'd be great.

499

Yeah, I think so for the first section, the first two questions, the candidate nailed it out of the park.

500

The only thing there was that he used a left join where I think we could have used an inner join, which is more efficient generally.

501

The other thing high level to note is that with these tables,

502

I generally like it's helpful when the candidate pushes the interviewer to ask about whether

503

or not the rows for each table are unique.

504

Because that's how the SQL should be written out.

505

I think in the understanding the table section of that section, I guess, like just general tip,

506

pushing the interviewer to give you as much information about the tables, you know, like, and unique, whether or not the row, the ID is unique, you know,

507

making these, making it really explicit what assumptions you're making there.

508

is really helpful because at the end of the day, if you're a SQL practitioner, you know that like all of the bugs that you have,

509

not all of it, like 80% is probably in the joins and these kinds of places.

510

So just knowing to ask those questions is a good, in my experience, I think that's a good way to conduct those interviews.

511

The third question was not very well worded in my opinion.

512

So that's a doc on the interviewer, I would say.

513

But it was meant to get at the window function, like the rank window function.

514

And I think the candidate did a good job of articulating how that should be written out.

515

But we ran out of time to be able to get to that.

516

So part of that is I'm not sure if there were too many questions for the amount of time we've allotted.

517

So, yeah, I'm not sure about that.

518

Did it feel realistic?

519

Khalid, did the interview, just in terms of surface area covered, did it feel realistic for you?

520

Yeah, it did.

521

And something that you mentioned, I think, on the first one, I was also hesitating to, I didn't know if I should bring up,

522

I think the candidate was focused a lot more on solving the problem.

523

And it's hard sometimes to like, hey, what sort of how do you engage the interviewer

524

while you're building the SQL query and you're kind of like thinking your thoughts aloud

525

when you mentioned unique values another question

526

that popped into my head was should I bring up missing values

527

or null values how are we going to track

528

that in these queries yeah so that's probably another question that could have been asked Yeah, that's really good.

529

I think that would, yeah, those are generally, in my experience, pretty helpful.

530

Yeah, moving on to the stat section, I thought overall it was pretty good.

531

Khalid said that this was the weakest section.

532

Was this Khalid's weakest section?

533

Was this the candidate's weakest section?

534

Yeah.

535

I would say if I had to choose, yes, because I think the first section was strong and the last section was strong.

536

Okay that um here i think um the candidate did a great job of explaining the central limit theorem um

537

at the first question um and

538

and i think also like i i had strong signal

539

or i guess the interviewer had a strong signal um on um whether

540

or not the candidate would be able to implement like a test to be able to compare and analyze the results.

541

So I didn't really have any questions there.

542

I think where I found a little bit of a gap was

543

when we started to talk about maybe the more intuitive explanation for statistical significance.

544

And that's relevant because generally,

545

like, I think more experienced data scientists will have to interface with non-technical stakeholders quite a bit and so

546

that actually is is a pretty relevant skill set to test for in an interview setting like this

547

and so um being able to articulate like hey it's

548

because intuitively you know there's this distribution we assume about these two groups

549

and if they're both coming from the same group then the difference would be like not statistically significant

550

but if they're actually coming from two different distributions then it is statistically

551

and that's what this test is measuring so that they can see that, oh yeah, like this difference is a real difference.

552

You know, that, that kind of explanation is what marketers would be able to, you know, understand a little better.

553

So that, that I wouldn't necessarily dock a candidate for that one, but I think, I think it's like a seniority level,

554

like experience level signal, I would say.

555

Just as your own experience as an interviewer Khalid, you probably have done these kinds of of stats like questions

556

before do you can you can you um can you speak a little bit to this point

557

that johnny just made about like that kind of technical

558

that kind of communication to non-technical stakeholders

559

that can make it make sense for them can you speak to that is

560

that in your mind a seniority signal as well uh yeah

561

i i do want to like be up front here

562

that i am i have done data science kind of early on in my career

563

and then have swapped into full stack and so I'm kind of now coming back in.

564

Cool.

565

Which is why that maybe you might understand in the interview that I kind of

566

still remember the high level ideas but I was at that level of time and the position that I was at

567

I was doing a lot more implementation.

568

And so it shows even in the interview, as you can see, where I get what I need to use and how to do it,

569

but not so much.

570

I seem to be missing some of that kind of like foundational stuff of, okay, how do you break down this idea for other stakeholders?

571

It's okay.

572

I mean, honestly, Johnny was like super accurate.

573

He's like, yeah, I can tell he knows how to do it, but I'm not sure that he's done it before.

574

Yeah, but everyone can relate to this.

575

That's doing a career pivot or is getting back into data science.

576

Do you two have any recommendations for people to strengthen that skill, whether they're rusty and getting back into DS or just like pivoting into DS as a career?

577

I feel like there's such a difficult balance right now.

578

And the one thing I'd love to talk about is like the roles, right? as an applied ai

579

or applied ml you don't really need to know some of this foundational level stuff i would argue

580

that you at all you need to know is

581

that i'm making an api call to chat gbt or an agent

582

and that'll help you get the job done whereas something like

583

a data science role is um maybe johnny can speak more

584

to this as a data science role you really have to understand some of the statistics

585

and the math part in order to make a better decisions

586

and really be able to do the job properly would you would you agree with

587

that johnny yeah i totally would

588

and um yeah i i think um for the most part

589

like it depends on the kind of data science role as

590

well um data scientist is like kind of a broad term in my experience like company to company

591

and it's usually some mix of

592

the analysis slash almost like consultant like product manager like kind of type of person

593

and then the engineer type person who's like implementing stuff

594

and like actually coding or whatnot

595

and the the scales of like what's more like prominent in

596

the role it's kind of differs company to company um where

597

i'm at right now at um at meta it's it's very much like these communication skills are very important

598

and it um they expect the data scientists to kind of

599

stand on their own feet in terms of being able to drive conversations

600

and influence and things things like

601

that um other roles it's more like your your engineering work is kind of where where um they're focused on

602

so i i really think it depends there um i think i just had a like having said

603

that i I think being able to communicate to non-stakeholders generally

604

is like a seniority experience level kind of skill set to have.

605

So it's always good.

606

But like when you need that, like how important is that for the hire specifically?

607

Like I don't that's why I said it's more like a seniority signal than a hire versus not hire, I think.

608

Awesome.

609

Awesome.

610

Really, really cool.

611

What about that last section?

612

The business case?

613

Johnny, what are your what are your notes here?

614

yeah um i thought it was really good um on on the whole

615

if i had to stack rank just for clarity i would say the school section was the best

616

and then this was the second and then the stats section on

617

i think for this section what i generally look for is structured thinking

618

and ability to start out with kind of like an exploration level

619

and then hone into hypotheses and as we go down be able to actually develop the metrics and maybe even a test,

620

design a test to validate those hypotheses.

621

So on the whole, I think that the candidate did it really well, I think.

622

But here also, like a measure of deeper experience would be

623

if the candidate can go end to end with that on their own and kind of drive the conversation forward.

624

I think the interviewer had to do a little bit of

625

kind of through the questions kind of shaping the answers along the way.

626

So I guess what I mean by that is, so I think up till like what hypothesis do you have for this drop,

627

like that question, like it's okay.

628

But then from there on being able to move to hypotheses that make sense.

629

And then like kind of going deeper and kind of driving the whole flow of the analysis in the conversation,

630

that's something I would look for as an interviewer, I guess.

631

And so I think my experience was that when I asked a question, there were answers that were pretty solid.

632

But I think

633

if there was a little bit more of being able to present the whole picture in a fuller way

634

and to kind of drive and kind of, you know, maybe stop at checkpoints to check with the interviewer to see if this is the right direction to go into.

635

That's something that generally shows like comfort level with the structured thinking

636

and experience having done an analysis of these kinds of things, of these types.

637

Amazing.

638

Amazing analysis.

639

Before we get Johnny's decision, any other things that either of you wanted to say point out about what was interesting or cool

640

or unexpected about this interview uh i i didn't realize how much stats would again this is like coming from someone who's

641

getting back into data science i didn't realize the stats portion

642

was very intriguing to me it's like wow this is a

643

lot of statistics um kind of my thoughts sort of my

644

expectation coming in was like oh this was going to be a design a recommendation system for google

645

and you go start from the beginning of the pipeline you do the some of the high level building blocks

646

and then you dive deeper but the reality is is

647

that it seems to be a little bit more targeted on the statistics especially and then the metrics portion

648

that was I think one of my biggest notice the other

649

one really interesting to hear Johnny say like I should be the one driving

650

or the candidate rather should be the one driving the conversation

651

right that's definitely a tip that I'm going to keep in the back of my head

652

because I guess my mindset was definitely here are the options it's up to you as the stakeholder to decide

653

which one of these paths you think we should go down.

654

So that's definitely a tweak that I'll probably take with me.

655

Yeah.

656

And I just wanted to jump in to say, I think like to your first point, that yeah, it might be

657

that this is more of a certain type of data scientist interview

658

that we're doing right now relative to the maybe more AI engineer type roles,

659

it sounds like, that you might be maybe perhaps aiming for as well.

660

So yeah, I would take that with a grain of salt, like in that, like it might not be as stats heavy as,

661

you know, some of those AI engineering roles, for example, might not be as stats heavy as like a data scientist role.

662

So I don't really know, like a lot of those roles are like brand new too.

663

So yeah, that's something I just want to look at.

664

All right.

665

And Johnny, your decision for this candidate?

666

Yeah, I think for like an entry to mid-level, definitely a higher.

667

I think at the senior level, I think there were some things that I saw as gaps.

668

But overall, I would have been happy to have somebody like this on the team.

669

All right.

670

Thank you.

671

That's all we got.

672

Well done.

673

And thank you for watching.

674

you

📺 Same Channel

✨ Recommend Video

260 sentences

Data Analyst Interview 2024 | SQL | POWER BI | Data Science | Mock Interview

University StudentsBusiness English+2

405 sentences

Netflix product manager mock interview and tips (with ex-Netflix Lead PM)

University StudentsBusiness English+2

154 sentences

Manual Testing: QA Interview Preparation Questions and Answers

University StudentsBusiness English+2

62 sentences

Day In my Life as a Software Engineering Intern at Google

University StudentsBusiness English+2

257 sentences

A/B Testing in Data Science Interviews by a Google Data Scientist | DataInterview

University StudentsBusiness English+2

370 sentences

FullStack Interview Questions (Junior & Mid)

University StudentsBusiness English+2

160 sentences

[2026] TOEFL iBT Speaking Take an Interview | Step-by-Step Strategies #0008

TOEFL PrepUniversity Students+3

324 sentences

Distributed Task Scheduler: System Design Interview (Stripe & Amazon Offers)

University StudentsBusiness English+2

204 sentences

Prompt Engineering - Guidelines

University StudentsBusiness English+1

278 sentences

Fastest way to become an AI Engineer in 2026

University StudentsBusiness English+2

22 sentences

I'm an ex-Google interviewer. You're doing LeetCode wrong.

University StudentsBusiness English+2

What is the Shadowing Technique?

Shadowing is a science-backed language learning technique originally developed for professional interpreter training and popularized by polyglot Dr. Alexander Arguelles. The method is simple but powerful: you listen to native English audio and immediately repeat it out loud — like a shadow following the speaker with just a 1–2 second delay. Unlike passive listening or grammar drills, shadowing forces your brain and mouth muscles to simultaneously process and reproduce real speech patterns. Research shows it significantly improves pronunciation accuracy, intonation, rhythm, connected speech, listening comprehension, and speaking fluency — making it one of the most effective methods for IELTS Speaking preparation and real-world English communication.

Shadowing technique: read the full step-by-step guide →

Shadowing Practice: FULL Data Science Mock Interview | SQL, Statistics, Business - Learn English Speaking with Video

Uber Data Engineering Mock Interview - Ride-Sharing Data Warehouse Schema

What is an API?

System Design Mock Interview: Design Facebook Messenger

How to Answer System Design Interview Questions (Complete Guide)

System Design Mock Interview: Design WhatsApp

Senior Product Designer Mock Interview: App Critique of Yelp

Understanding TCP vs UDP

What REALLY Happens in Meta’s AI Product Sense Round

Anthropic Culture Fit | Behavioral Interview

Data Analyst Interview 2024 | SQL | POWER BI | Data Science | Mock Interview

Netflix product manager mock interview and tips (with ex-Netflix Lead PM)

Manual Testing: QA Interview Preparation Questions and Answers

Day In my Life as a Software Engineering Intern at Google

A/B Testing in Data Science Interviews by a Google Data Scientist | DataInterview

FullStack Interview Questions (Junior & Mid)

[2026] TOEFL iBT Speaking Take an Interview | Step-by-Step Strategies #0008

Distributed Task Scheduler: System Design Interview (Stripe & Amazon Offers)

Prompt Engineering - Guidelines

Fastest way to become an AI Engineer in 2026

I'm an ex-Google interviewer. You're doing LeetCode wrong.

Why Practice Speaking with This Video?

Grammar & Expressions in Context

Common Pronunciation Traps

What is the Shadowing Technique?